U.S. SUPERCOMPUTER VULNERABILITY A REPORT PREPARED BY: SCIENTIFIC SUPERCOMPUTER SUBCOMMITTEE COMMITTEE ON COMMUNICATIONS AND INFORMATION POLICY U.S. ACTIVITIES BOARD IEEE, INC., AUGUST 8, 1988
Document Type:
Collection:
Document Number (FOIA) /ESDN (CREST):
CIA-RDP90G01353R001100170002-4
Release Decision:
RIPPUB
Original Classification:
K
Document Page Count:
54
Document Creation Date:
December 23, 2016
Document Release Date:
August 16, 2012
Sequence Number:
2
Case Number:
Publication Date:
September 8, 1988
Content Type:
MEMO
File:
Attachment | Size |
---|---|
![]() | 2.82 MB |
Body:
STAT
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4
I
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4 FIVE SECRETARIAT / 1
TO.
ROUTING SLIP
ACTION
INFO
DATE
INITIAL
1
DCI
2
DDCI
X
-
3
EXDIR
4
D/ICS
5
DDI
X
6
DDA
7
DDO
8
DDS&T
9
Chm/NIC
10
GC
11
IG
12
Compt
13
D/OCA
14
D/PAO
15
D/PERS
16
D/Ex Staff
17
C/TTAC/OSWR
X
18
ZWCGI/DI
19
NIO/ECON
20
21
22
SUSPENSE
ER 570X
3637
mxuti eutary
14 SEP 88
Date
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4
STAT
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4
R
Next 1 Page(s) In Document Denied
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16:
CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
IEEE/USAB COMMITTEE ON COMMUNICATIONS AND INFORMATION POLICY
UNITED STATES ACTIVITIES BOARD
John M. Richardson
Chairman
(202) 334-2844
Cloud M. Davis
Vice Chairman
(914) 742-5929
Richard VanSlyke
Vice Chairman
(718) 260-3050
Heidi F. James
Executive Secretary
(202) 785-0017
PLEASE REPLY TO:
1111 19th Street, NW
Suite 608
Washington, DC 20036-3690 USA
U.S. SUPERCOMPUTER VULNERABILITY
A REPORT PREPARED BY:
SCIENTIFIC SUPERCOMPUTER SUBCOMMITTEE
COMMITTEE ON COMMUNICATIONS AND INFORMATION POLICY
UNITED STATES ACTIVITIES BOARD
INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, INC.
AUGUST 8, 1988
THE INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, INC.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
U.S. SUPERCOMPUTER VULNERABILITY
BY THE
IEEE SCIENTIFIC SUPERCOMPUTER SUBCOMMITTEE 1
COMMITTEE ON COMMUNICATIONS AND INFORMATION POLICY
INTRODUCTION
The U.S. created the supercomputer industry. Today U.S. firms jointly are
leaders in world supercomputer markets. Nonetheless, U.S. supercomputer firms
are vulnerable to a focused strategy by the Japanese targeted on their industry.
As documented below, the reasons for the vulnerability of the U.S. firms are
complex. To overcome their vulnerability will require a systems solution: an
integrated, cooperative effort among industry, universities and government. It
appears that such a solution will require coordination by government to a degree
seldom if ever achieved in this country in peace time. But the threat is real,
and it is not limited to supercomputers. Supercomputers appear to represent
only a next step in an on-going process that is gaining momentum. As one can
see from the "Visions" for the future promulgated by the Ministry of
International Trade and Industry (MITI), Japan seeks a dominant position in the
Information industries--the computer and communications industries of the world.
And it is well on the way to achieving that position.
Many doubt that the U.S. government could play an effective, leadership, policy
role in relation to civilian technology/industry problems of this kind. Yet the
Japanese government clearly does play such a role in its economy. And Japan is
a market-oriented, capitalistic, democracy. By using its approach Japan is
overtaking, and often surpassing, the U.S. in field after field. How long can
we afford to wait before we respond?
Perhaps the most important question for the U.S. to ponder is: What is the
alternative to effective government leadership and support in vital
technology/industry matters such as the ones discussed here? Current approaches
are not working. The discussion that follows is focused on the supercomputer
issue as compounded by the semiconductor issue. Yet, these two issues appear to
be only special cases of the broader, generic problem.
1 The IEEE Scientific Supercomputer Subcommittee is a technology-policy subcom-
mittee of the Committee on Communications and Information Policy under the
United States Activities Board of the Institute of Electrical and Electronics
Engineers, 1111 19th Street, NW, Washington, DC 20036. The members of the -
Subcommittee are:
Sidney Fernbach, Chairman
Lara Baker
Vito Bongiorno
Alfred E. Brenner
James F. Decker
Duncan Lawrie
Alan McAdams
Kenneth W. Neves
John Ranelletti
John P. Riganati
John M. Richardson, Chairman, Committee
on Communications and Information Policy
Stewart Saphier
Paul B. Schneck
Lloyd Thorndyke
Kenneth F. Tiede
Hugh Walsh
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declasiified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-2-
ANALYSIS
Cray and Control Data Corporation (CDC) with its ETA subsidiary are the two
long-term U.S. manufacturers of supercomputers. Cray has the largest world
market share. It is a stand-alone company, solely dependent on the supercom-
puter business. CDC and ETA are emerging from financial difficulties and, while
more diversified, are not yet nearly as strong as Cray in the supercomputer
field. Both Cray and ETA are in a vulnerable position.
Several important factors contribute to the vulnerability of the U.S. supercom-
puter firms: (1) the current situation in the semiconductor industry and the
recent history of US-Japan trade problems in the industry; (2) the rapid strides
the Japanese supercomputer firms have made; (3) the fact that Cray and ETA rely
on high performance memory devices that now are available only from Japan; (4)
the implications of the recent supercomputer trade agreement with Japan; and (5)
the fact that U.S. firms are losing a primary market advantage, their differen-
tial access to the huge reservoir of applications software that currently is
optimized primarily for their installed base of supercomputers.
These factors are discussed in relation to their specific impact on U.S. super-
computer manufacturers; but the process is generic Japanese. As stated in
SCIENCE in November, 1985:2
"...In this process, the Japanese begin by picking a few key,
high-volume components over which to compete. Then, as the
U.S. producers retreat under extreme price-cutting assaults,
the Japanese companies extend the fight step-by-step to other
items until they have swept competitors out of the most
profitable areas. (Emphasis added.)
This has happened in other industries, and in semiconductors
the Japanese have all but taken over the field of memory chips..."
Conditions are ripe for supercomputers to represent the next step in the pro-
cess.
RECENT SEMICONDUCTOR HISTORY
The U.S. created the semiconductor industry. The U.S. invented the VLSI (very
large scale integration) Chip. For decades U.S. firms jointly were the leaders
in world semiconductor markets. Nonetheless, U.S. semiconductor firms were
vulnerable to a focused strategy by the Japanese on DRAMs, dynamic random
access memories.
The Japanese developed and manufactured excellent, highly reliable, high quality
DRAMs. Then, (as adjudicated by the U.S. Foreign Trade Commission) for more
than two years the Japanese dumped these products (sold them below cost) in the
U.S. and third markets. During the time that the Japanese were flooding the
market with DRAMs, U.S. firms lost in excess of two billion dollars. U.S.
merchant manufacturers of semiconductors are stand-alone companies. Effectively
they were forced to withdraw from domestic production of high-performance DRAMs.
2 Eliot Marshall, "Fallout from the Trade War in Chips," SCIENCE,
November 22, 1985, p. 918.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declasiified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-3-
In other product lines, SRAMs, static random access memories; and EPROMs, era-
sable programmable read only memory devices; U.S. firms also experienced duress.
The U.S. firms responded with legal (as opposed to technological) actions that
led to the U.S. Government's imposing tariffs on the Japanese as sanctions.
This government effort, while well intended, raised the prices of chips and thus
resulted in increased costs to U.S. supercomputer manufacturers and thus has
tended to reduce their competitiveness.
More recently, the problem of dumping by the Japanese has been replaced by
the current problem of severe shortages of DRAMs in the U.S. These shortages
have been shown by experts from the Brookings Institution to result from reduced
actual levels of Japanese output that "happen" to accord almost perfectly with
the "forecasts" of production levels that had been made by MITI several months
before. DRAM prices in the U.S. have sky rocketed--sometimes tenfold, and more.
The DRAM business is now enormously profitable to the Japanese producers of
semiconductors. This is a classic pattern for those who achieve effective domi-
nance of a market.
During the period of DRAM shortages, exchange rates have been highly favorable
to U.S. domestic production. Furthermore, sales of DRAMs in the U.S. market
have been highly profitable to suppliers. Nonetheless, (at the time of this
writing) U.S. firms were not re-entering the market, either by investing in new
plants or even by reopening "mothballed" DRAM plants. Perhaps they recall their
recent huge losses too vividly, or perhaps the rapid "decay" of know-how and
expertise in so demanding a business explains their lack of response. In any
case, a market that is difficult to re-enter is the type of market in which the
classic pattern is most successful.
The problems of U.S. firms are rooted in the new-found power of Japanese firms
In important aspects of the semiconductor business. In presentations to the
authoring IEEE groups, expert observers have stated variously that Japan has
"won the silicon war," (August, 1985); and that, "We (the U.S.) failed. We had
it (semiconductor leadership) and we lost it!" (May, 1988). Despite these
pessimistic assessments and the current dire situation, we believe that this
need not (yet) be the case. But the hour is late, and strong action will be
required to prevent it.
JAPANESE SUPERCOMPUTER COMPETITION
Cray and CDC/ETA face the three Japanese manufacturers of supercomputers shown
on Table I. Not coincidentally, these three firms are also among the leading
Japanese manufacturers of semiconductors. Cray and ETA are tiny in comparison
to their Japanese rivals--they are small even in relation to the leading U.S.
merchant semiconductor manufacturers. (These firms are also shown on Table I.)
Given the experience of the semiconductor industry, plus the facts cited in the
next paragraphs, it is not hard to understand why the U.S. supercomputer firms
are vulnerable to the "next step" by the Japanese.
The sum of the current annual supercomputer outputs of the three Japanese giants
approximates the number, though not yet the dollar value, of Cray's output, and
greatly exceeds that of CDC/ETA. As Table I demonstrates, the Japanese firms
are not stand-alone manufacturers of supercomputers. Their operations are
integrated over multiple production stages and across several product lines.
Table II shows the number of supercomputers of each manufacturer installed and
on order as of late 1987. It is important to note that both the CDC 205 and the
Cray 1 are now obsolete, while the ETA 10's and Cray Y-MPs (not listed) are just
becoming operational. Table III shows the peak megaflops per processor for each
manufacturer as well as the megaflops per total (multiprocessor) system.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-4-
Table I
Market Segments Participated in By Selected U.S. And
Japanese Manufacturers of Supercomputers and/or Semiconductors
TI
INTEL
Motorola
AMD
CRAY CDC/ETA ETC. AT&T IBM NEC FUJITSU HITACHI
Super- x x RE* x
computers
Main Frame
Computers
Intermediate x x x x
Computers
Mini Computers x x x x x x
Micro Computers x x x
Consumer
Electronics x X x
Semiconductors
-Merchant x x x x
-Captive x x x x x
* Re-entering.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
beclasified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-5-
Table II
Supercomputer Installations
System
No.
No. on
Order
FCS
CDC/ETA
Installed
CDC Cyber 205
37
0
1980
ETA 10 series
2
8
1987
Cray
Cray-2
8
4
1985
X-MP series
117
43
1984
Cray-1
58
0
1976
Fujitsu
VP series
56
1984
Hitachi
S-810 & 820
19
4
1983
NEC
SX2 series
12
7
1985
Source: Adapted from The Gartner Group,
Table III
Peak Performance Rates
SINGLE CPU PEAK
COMPUTER 1 64-BIT MFLOP RATE 1
November, 1987.
ALLOWING MULTIPLE
CPUs
CRAY-1
CRAY X-MP
CRAY-2
CRAY-3 (1989)
CYBER 205
ETA-10 (1986)
ETA-10/E
ETA-10/G (1988)
FUJ. VP 100
FUJ. VP 200
FUJ. VP 400
HIT. S-810/20
HIT. S-820/80
IBM 3090/VF
NEC SX-1
NEC SX-2
NEC SX-3 (1989)
1
Source: Adapted from "Computational Fluid Dynamics: Algorithms and Supercom-
puters," by W. Gentzsch and K. Neves, NATO AGARDograph No. 311, March, 1988.
160
160
233
932
(4-CPUs)
488
1,952
(4-CPUs)
1,000
16,000
(16-CPUs)
est.
200
(2-PIPE)
400
(4-PIPE)
350
1,400
(4-CPUs)
415
1,660
(4-CPUs)
643
5,142
(8-CPUs)
271
271
533
533
1,067
1,067
630
630
2,000
N/A
116
696
(6-CPUs)
570
570
1,300
1,300
5,000
20,000
(4-CPUs)
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-6-
The facts in these tables are consistent with the following summary. The
Japanese manufacturers, especially Fujitsu, have made rapid progress in
installing systems during the four short years they have been producing super-
computers (Table II). Their systems are to date single processor systems (Table
III). This latter point relates to two elements: (A) the individual processors
In the Japanese supercomputers are more powerful today than any U.S. manufac-
turer currently expects to deliver in less than half a decade; but, (B) the
Japanese manufacturers today lag somewhat in adapting to parallel processing
techniques (though they appear to lead in the complementary field of vec-
torization and vector processing; see below).
These facts imply that multiprocessor versions of the Japanese supercomputers,
especially the newest Hitachi and NEC (Nippon Electric Corporation) machines,
could far outstrip the best of the U.S. machines once the Japanese catch up in
parallel processing, an area where they are hard at work. NEC has announced
parallelism for its SX-2 series and is said to have completed the circuit design
for its 20 gigaflop, four processor SX-3 series currently in development.
In a recent special issue (March 3, 1988) of the trade journal Electronics
entitled "Inside Technology,"3 these facts are reiterated in very strong
language:
"...The Japanese stick to simpler single-processor architectures.
And their supercomputers are the speediest one-processor machines
In the world--by far--thanks primarily to advanced semiconductor
technology. The only way U.S. supercomputers can come near their
processing rates is in multiprocessor configurations."
The following quotes also are in the article in a box entitled, "Where Japan
Shines: Vectorizing Compilers."4
"... Japanese compilers make it relatively easy for a user
with his own Fortran program to wring near maximum performance
from a supercomputer's vector processor without resorting to
special vector-processor programming methods.
Automatic vectorization has been a Japanese strong suit for some
time, and it is made easier by the Japanese's use of single-processor
architectures in supercomputers. Compiler writers need not face
the difficulties of automatic parallelization."
The latter quote points up an additional area of Japanese strength; the systems
software of the Japanese supercomputers is already world class. NEC's Fortran
compiler is reputed to be the best in the world. If that is in fact the case,
It would have surpassed the previous best--that from Hitachi.
3 Charles L. Cohen, "Japan Focuses on Simple but Fast Single-Processor
Supercomputers," Electronics, March 3, 1988, p. 57.
4 Ibid. p. 58.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-7-
U.S. SUPERCOMPUTER FIRMS RELIANCE ON JAPANESE SEMICONDUCTORS
Already the aggressive pricing strategies of the Japanese in supercomputers
(discounts of up to 80% to universities and others) combined with inversely
aggressive semiconductor strategies (refusal to export their latest high perfor-
mance component devices), threaten the existence of Cray and ETA. The highest
performance memory and bipolar logic components useful for supercomputers are no
longer manufactured in the U.S.; they are available only from Japan. The
managements of Cray and ETA have been quoted in the press at various times as
stating that these Japanese components are, "not yet available for export" from
Japan to Cray or ETA as devices--but they are available to end users in the
Japanese supercomputer systems. Those systems are definitely available for
export.
This continues the familiar, oft-repeated pattern. Japanese firms plan and act
In accord with long-range strategic goals. When they achieve a technological
advantage in one area, they use that advantage to ensure their advance into new
areas. They have targeted supercomputers as the next high tech area in which
to establish a dominant position, and they have made enormous strides toward
that goal as Tables II and III demonstrate.
THE IMPLICATIONS OF THE SUPERCOMPUTER TRADE AGREEMENT WITH JAPAN
The supercomputer trade agreement with Japan contains the seeds which can
further undermine the position of U.S. manufacturers. It requires of Japan a
number of actions which, if made reciprocal to the U.S., could facilitate the
entry of Japanese supercomputers into the U.S. market.
The reasoning that supports this conclusion proceeds through the following
steps. Almost by definition, no one knows how to use a new generation of super-
computers efficiently. New systems generally embody the latest technology and
architectural design in order to achieve their state-of-the-art computing
speeds. In this country, the national laboratories historically have been
"partners" in supercomputer development. The laboratories have developed the
applications and forced the evolution of the operating systems that together
permit the machines to work effectively. In other words, the laboratories'
expert users have assisted greatly in facilitating improved productivity of the
systems as these users learned more about the characteristics of the new systems
and how to make optimal use of them. Now with the new NSF programs that have
made supercomputers available to universities, universities, too, are becoming
partners in these efforts. Often the U.S. manufacturer has recognized the
"'partnership contribution" of these U.S. institutions through price concessions,
"buybacks" of time or other means. The same has been true in Japan.
The trade agreement requires the Japanese in procurements of supercomputers
by their government or their universities to give equal preference to U.S.
manufactured supercomputers. In essence, the Japanese have agreed to avoid
"unfair" pricing on the part of the Japanese manufacturers. But what may seem
"unfair" to a trade administrator looks to a manufacturer like "recognition
of partnership" with a government agency or university.
While negotiating this agreement with one hand, the U.S. government has blocked
the sale of a NEC supercomputer to M.I.T. with the other. To date Japanese
supercomputers have been virtually shut out of U.S. government and university
markets. It appears that government agencies, especially the Department of
Defense, intend to keep it that way.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-8-
How long can such a pattern last? Is it realistic to believe that the Japanese
can long be required to give unbiased consideration to U.S. supercomputers while
Japanese machines are foreclosed from U.S. institutions? If the requirements we
now impose on the Japanese were made reciprocal for U.S. universities and
government laboratories, Japanese supercomputers would have to be acceptable to
those agencies on non-discriminatory terms. The Japanese systems are already a
match for the U.S. systems and soon could surpass them in performance. Price
thus is likely to be a major determinant of choice; the Japanese manufacturers
have been more than willing to sacrifice short-term profit for long-term market
penetration.
The new requirements could greatly facilitate the entry of the Japanese into
the U.S. At the same time such requirements could disrupt the implicit partner-
ship between U.S. manufacturers and U.S. government laboratories and/or U.S.
universities, and transfer the benefits of partnership to Japanese firms.
In response to a low bid, U.S. National Laboratories could be required to become
partners to Japanese firms in perfecting their systems for penetration of U.S.
markets.
INCREASING "PORTABILITY" OF SOFTWARE
Paradoxically, desirable developments long sought by users on another front are
hastening the progress of the Japanese. Itis becoming increasingly easy to
move applications from the machine of one supercomputer vendor to that of
another. There are four elements which are facilitating software "portability:"
(1) Standardized versions of FORTRAN;
(2) Improved FORTRAN optimizing compilers;
(3) Broad-spectrum libraries of algorithms optimized for each
vendor's systems;
(4) A common operating system--UNIX--now being implemented on supercomputers.
While in the past each supercomputer vendor had a "captive audience" of users,
now it is increasingly possible to move applications between and among vendors.
As the industry leader with the largest installed base, Cray had benefited from
the prior situation in the industry; CDC did also, but to a lesser extent. But
portable software does provide substantial benefit to users of supercomputers,
it greatly diminishes this advantage previously enjoyed by the U.S. firms.
New entry into the field of supercomputing also is facilitated by increased
applications portability. As the most recent entrants, the Japanese vendors are
the greatest beneficiaries of these changes.
Cray itself also benefits from portability since portability helps Cray provide
compatibility to its own multiple product lines. Use of UNIX permits Cray to
focus on a single operating system thus reducing its development costs. The
most significant effects, however, are those on the new entrants.
A further benefit to users is also achieved. Through UNIX, an operating
environment uniform from work station to minicomputer, to minisupercomputer, to
supercomputer would become possible. The Los Alamos National Laboratory is one
of the organizations working hard to bring this about. One of its major objec-
tives is to bring the myriad, powerful software innovations pioneered in the
highly competitive fields of commercial computing over into the more rarified
arena of supercomputing--with benefit to U.S. manufacturers as well as to users.
The other side of the coin, of course, is that these benefits accrue to all ven-
dors of supercomputers--especially to the Japanese.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-9-
THE IMPLICATIONS OF THE PHENOMENA OF SEMICONDUCTOR DISADVANTAGE AND THE
PRESENCE OF PORTABILITY
At this point two threads come together. Once general portability-of applica-
tions is achieved, the avenue to continued market leadership by U.S. firms would
be solely through technological leadership. But how can U.S. firms simulta-
neously rely on componentry manufactured by their competitors, the Japanese, and
assure their customers that they, the U.S. firms, can maintain technological
leadership? The Japanese dominate completely in semiconductor memory (and in
bi-polar logic). The American firms have each sought logic components of usual
technology to advance their competitive positions: Cray with Gallium Arsenide
chips for the Cray III, ETA with ultra-dense C-MOS chips cooled with liquid
nitrogen for the ETA-10. (The latter chips are available from Honeywell as a
result of that firm's participation in the VHSIC "Very High Speed Integrated
Circuit" program of the Department of Defense. To date, these have not been
manufactured to the specifications required for ETA to meet its advertised
cycle-time of 7 nanoseconds.) But we see also that despite these efforts, at
their rated performance levels the U.S. systems lose the megaflops/processor
race (Table III). And each firm's effort is based on a favorable "spike" in an
otherwise dismal U.S. semiconductor picture--which as yet shows no signs of
Improvement sufficient to reverse its overall, rapid decline.
From experience we can see that once the Japanese fully consolidate their
semiconductor technology leadership, in time U.S. firms would be most fortunate
if they found themselves to be only a generation or so behind them. The
experience of Asia's "Four Tigers," Korea, Taiwan, Hong Kong, and Singapore con-
firms this point. The Japanese have consistently resisted the requests of the
tigers for newer Japanese technologies. The tigers are explicitly told by the
Japanese that this is because Japan looks upon them as "potential future
competitors." As a rule, the Japanese refuse to license them for'any technology
less than five years old.
LONGER RUN IMPLICATIONS
The "vector facility" which provides vector capability to IBM's 3090 Mainframes,
today blurs the boundary between the mainframe and the supercomputer. In any
case, IBM has demonstrated its renewed interest in the field of supercom-
puting. This confirms the field's growing commercial importance--but IBM is not
yet well established as a vendor of supercomputers.
The Japanese are established in supercomputing. They are challenging IBM in
mainframes: Hitachi, with its plug compatible CPU's marketed by National
Semiconductor's subsidiary, NAS; Fujitsu/Amdahl with their CPU's marketed
through Amdahl Inc. (which recently experienced its best year). NEC's ACOS
series is marketed in the U.S. through Honeywell. The mainframes from all three
of these vendors also have some vector capability. Most Japanese supercomputers
are "IBM-compatible." This is especially true of those from Fujitsu and
Hitachi. (And the Fortran for NEC's SX-2 series is known as its "IBM version."
For the SX-3 series NEC plans both an "IBM version" and a "Cray version.")
Mainframes are growing larger and faster. But the Japanese appear to have
Intercepted that market from above through their supercomputers.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-10-
CONCLUSION
Once they were to achieve preeminence in supercomputers, the Japanese would
have established themselves as the manufacturers of the most powerful computer
systems, with supercomputers installed with the "Gold Chip" accounts of the
world. They could then use that base and prestige to cement their move into the
more lucrative market for commercial data processing systems. That is a long-
term goal.
The threat is not just to supercomputers. The threat is to the entire U.S.
base of computer systems. It is ironic that such cascading impacts could
originate from dominance of a product area--DRAMs--that has reached almost com-
modity status. Yet such impacts appear not only possible, but likely. There is
no denying the enormous impact that current DRAM shortages are already having on
U.S. systems manufacturers. Their component costs are rising sharply. Many are
unable to meet their orders due solely to lack of DRAMs.
The most recent developments with Sematech further confirm the depth of the cri-
sis in U.S. semiconductor manufacturing at two levels. First, IBM and AT&T each
has agreed to donate an advanced technology to Sematech. For IBM it is the four
megabit DRAM. For AT&T it is the technologically equivalent one megabit SRAM.
That these two arch rivals would see their way clear to share technology with
each other and with others in the U.S. semiconductor industry demonstrates the
depth of their concern. In part this is because these integrated firms buy
their equipment for manufacturing semiconductors from others. And the Japanese
are dangerously close to becoming the sole source of that equipment. Second, it
is not encouraging that (as of this writing) this fledgling effort at coopera-
tion by the semiconductor industry, Sematech, remains understaffed, organiza-
tionally in disarray, and well behind its self-imposed schedule.
At another level, it must be recognized that there is no assurance that even a
Sematech highly successful in relation to its stated goals was set up to be
coordinated with needed actions in relation to supercomputers--or with any other
elements required for an effective national technology policy. This brings us
back to the point made in the opening paragraph.
REQUIRED: A SYSTEM SOLUTION
The solution to the problems of the U.S. supercomputer manufacturers lies not in
imploring the Japanese not to pursue their advantage, but in the U.S. taking
positive actions at home to insure that the Japanese don't succeed at our
expense. The U.S. must get its technology base in order: in semiconductors, in
supercomputers and in software. No current initiative--including Sematech or
IBM's own recent focus on the field of supercomputers--assures that these objec-
tives will be achieved. A system solution cannot be achieved through discrete,
uncoordinated initiatives, no matter how worthy.
Several studies over recent years make ft clear that a solution to the problems
of U.S. civilian technology/industry/trade policy, such as the matters discussed
in this paper, will require expert coordination and leadership.5 It seems clear
5 For example, see "Global Competition: The New Reality," the Report of the
President's Commission on Industrial Competitiveness, January, 1985;
Superintendent of Documents, Washington, DC, 20402; and "The Technological
Dimensions of International Competitiveness," May, 1988, Office of
Administration, Finance, and Public Awareness, National Academy of Engineering,
2101 Constitution Ave, NW, Washington, DC 20418.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
beclassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-11-
that to be effective, these can only by provided at the national level. Only at
this level would it be possible for a true systems approach to be taken to the
problems facing this country in the international competitiveness of its
Industry and of its high technology base. Action is required on long term R&D
policies: it is necessary to integrate the activities of government, industry
and universities; it is necessary to achieve timely, effective technology
transfer from stage to stage in the process; it's necessary to ensure careful
attention to the provision of suitable people-power to each stage in the pro-
cess, an issue involving not just numbers, but also appropriate intellectual
capital.
The challenge is to find an acceptable institutional framework in which govern-
ment, industry and academia can pursue these objectives to the long-run benefit
of the nation as a whole. A necessary requirement of such a framework is: it
must ensure that economic and technological decisions be taken on the basis of
economic and technological--as opposed to mainly political or military--
criteria. The answer may well require that the coordination and leadership
functions be vested in a new, lean, expert, civilian agency of government that
is capable of focusing on the longer term national interest. Only through a
coordinated approach to all of the above issues will we be able to ensure a
strong U.S. base for innovation, productivity, and international com-
petitiveness--a base in which supercomputers constitute a vital factor.
For more information on the IEEE Scientific Supercomputer Subcommittee, contact:
Heidi F. James
IEEE/USAB
1111 19th Street, NW
Suite 608
Washington, DC 20036
(202) 785-0017
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized 'Copy Approved for Release 2012/08/16: CIA-RDP90G01353R6-6-1160170002-4
IEEE
E IiNSTITUTE. OF ELECTRICAL ANti kLEcTRONIcs ENGINEERS, INC. ? 4,
FOR IMMEDIATE RELEASE
Contact: Pender M. McCarter
202/785-0017
or
Jayne F. Cerone
212/705-7847
VARIOUS JAPANESE TECHNICAL, MARKETING STRATEGIES
MAKE U.S. SUPERCOMPUTER FIRMS VULNERABLE:
IEEE SUPERCOMPUTER SUBCOMMITTEE REPORT
U.S. Lead in Supercomputing Could Be Lost;
IEEE Group Urges Focus on 'Longer-Term National Interest'
to Ensure Strong Technology Base
WASHINGTON, DC, August 8: Various Japanese technical and marketing
strategies make U.S. supercomputer firms "vulnerable to loss of their
world leadership," according to a report issued today by the Scientific
Supercomputer Subcommittee of the IEEE Committee on Communications and
Information Policy (CCIP). The report, titled "U.S. Supercomputer
Vulnerability," was prepared by the CCIP of The Institute of Electrical
and Electronics Engineers, Inc. (IEEE). It cites Japan's introduction of
advanced machines and adoption of aggressive marketing techniques,
including use of strategic delays in marketing high-speed computer chips
in the United States. And the report recommended that the U.S. focus on
the longer-term national interest to ensure a strong technology base.
If allowed to become preeminent in supercomputers, Japanese computer
companies could then use that base and prestige to increase their role
in the more lucrative market for processing systems, says the IEEE
-more-
NEW YORK, N.Y. ? 345 EAST 47TH STREET, 10017
WASHINGTON, D. C. ? 1111 NINETEENTH STREET, NW. 20036
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-2-
group of academic, government, and industry experts. It also observes
that these companies include operations integrated over multiple pro-
duction stages and across several product lines.
The report adds that, although the most advanced individual processors
in Japanese supercomputers are superior to those used in U.S.
machines, the Japanese still lag somewhat in their ability to perform
parallel processing; that is, solving problems faster by dividing them
into parts that can be handled by a number of processors simulta-
neously. However, new multiprocessor versions of Japanese machines
"could far outstrip the best of the U.S. machines once the Japanese
catch up in parallel processing, an area where they are hard at work,"
the report states.
Paradoxically, the increased "portability" of programs among supercom-
puters of different manufacturers, aided by a standardized FORTRAN
computer language and a common UNIX operating system -- otherwise
desirable, is also "hastening the progress of the Japanese," according
to the IEEE group. It summarizes: "While in the past each [U.S.]
supercomputer vendor had a 'captive audience' of users, now it is
increasingly possible to move applications between and among vendors.
Portable software does provide substantial benefit to users of super-
computers, but it greatly diminishes the advantage previously enjoyed
by U.S. firms." In addition, the report says, Japanese dominance
in producing high-performance chips could mean that, "in time, U.S.
firms would be most fortunate if they found themselves . . . [only to
be] a generation or so behind [the Japanese]."
-more-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
-3-
The document also addresses the recent U.S. supercomputer trade
agreement with Japan. It states that "Japanese manufacturers have been
more than willing to sacrifice short-term profit for long-term market
penetration." In this way, the report continues, the trade agreement
could work to the disadvantage of the U.S. as Japanese supercomputers
become more competitive with U.S. machines. The document also notes
the seriousness of the Japanese challenge, pointing to a slow start
for Sematech, the industry consortium established to encourage tech-
nological advances in the semiconductor industry.
The IEEE-CCIP subcommittee stresses that "the solution to the problems
of . . . U.S. supercomputer manufacturers lies not in imploring the
Japanese not to pursue their advantage, but in the U.S. taking posi-
tive actions at home to ensure that the Japanese don't succeed at our
expense." The group calls for the creation of Federal research and
development policies that will integrate efforts of government,
Industry, and universities -- all focusing on the longer-term national
Interest to ensure a strong U.S. base in supercomputer and other
critical technologies.
The IEEE-CCIP is chaired by Dr. John M. Richardson of the National
Research Council. CCIP's Scientific Supercomputer Subcommittee is
headed by Dr. Sidney Fernbach, a pioneer supercomputer user, former
director of the computer center at the Lawrence Livermore National
Laboratory, and now a consultant to Control Data Corporation.
# # #
[Note to Editors: Copies of the report are available from Pender M.
McCarter, Manager, Public Relations, IEEE Washington Office, telephone
(202) 785-0017; and Jayne F. Cerone, Coordinator, Media Relations,
IEEE Headquarters, telephone (212) 705-8747.]
8/8/88
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
?Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
News in Perspective
SUPERCOMPUTING
IEEE Warns of the Japanese
Supercomputer Threat
A new study finds that the U.S. must act now to stem the
perceived Japanese invasion and suggests anew civilian
agency focused on long-term national interests.
BY WILLIE SCHATZ
The IEEE's Committee on
Communications and Infor-
mation Policy has made avail-
able to DATAMATION a new re-
port on how the Japanese are
wiping out the U.S. super-
computer industry. Entitled
"U.S. Supercomputer Vul-
nerability," it leads to the in-
exorable conclusion that "we
better do something," accord-
ing to its principal author,
Alan McAdams, an assistant
professor of managerial eco-
nomics at Cornell University.
The Scientific Super-
computer Subcommittee of
IEEE contends that the U.S.
supercomputer industry is in
deep trouble thanks to a fo-
cused market strategy by the
Japanese. To overcome U.S.
firms' vulnerability will re-
quire coordination by govern-
ment to a degree seldom
achieved in peacetime?and
time is of the essence.
What's new about this?
"The IEEE has never taken a
position like this before," Mc-
Adams says. "This is the non-
partisan IEEE taking a policy
position. Things must be
pretty bad for them to start
screaming."
Finding the Framework
The something Mc-
Adams refers to is finding
"an acceptable institutional
framework in which govern-
ment, industry, and academia
can pursue these objectives
to the long-run benefit of the
nation as a whole." Were that
to occur, the institutional
framework would still be use-
less without a guarantee that
economic and technological
decisions be made according
to economic and technologi-
cal?not military and politi-
cal?criteria.
"The answer may well
require that the coordination
and leadership functions be
nested in a new, lean, expert
civilian agency of government
that is capable of focusing on
the longer-term national in-
terest," find McAdams and
friends. "Only through a coor-
dinated approach to all these
Hopefully, the new adminis-
tration will have more of a
commitment to understand-
ing high tech. But it's still not
worth creating another
bureaucracy."
That opinion isn't con-
fined to the government.
"Any proposal to establish a
new government agency isn't
something I favor," says Sid
Karin, director of the National
Science Foundation's (NsF's)
San Diego Supercomputer
CORNELL'S McADAMS: The threat isn't limited to supercomputers.
issues will we able to ensure a
strong U.S. base for innova-
tion, productivity, and inter-
national competitiveness."
Grass-Roots Commitment
"A new agency is fine if
you've got a professional gov-
ernment with a long-term in-
terest," says a government
official intimately involved in
the supercomputer industry.
"But if Congress thinks it can
just create it with a few pieces
of paper, then it will be anoth-
er bureaucracy that won't
work. What you really need is
a grass-roots commitment.
Center (sDsc). "I've heard all
this before. There's nothing
earthshaking in here."
The IEEE begs to differ.
"People aren't realizing the
real crisis that exists," Mc-
Adams contends. "The whole
U.S. technological base is at
risk, and it's getting worse.
To emphasize that this
report is different from all the
others that have reached the
same conclusion, McAdams
rests his case on economics
and technology.
"To overcome [U.S. su-
percomputers] vulnerability
will require a systems solu-
tion: an integrated coopera-
tive effort among industry,
universities, and govern-
ment," the report says. "It ap-
pears that such a solution will
require coordination by gov-
ernment to a degree seldom if
ever achieved in this country
in peacetime. But the threat is
real, and it is not limited to
supercomputers. Supercom-
puters appear to represent
only a next step in an ongoing
process."
Banging the Drum Slowly
Thus, while Cray and
ETA may say that Japanese
components are not yet avail-
able for export to those two
companies, the devices are
readily available to end users
in the Japanese supercom-
puter systems.
"This continues the fa-
miliar, oft-repeated pattern,"
the report contends. "Japa-
nese firms plan and act in ac-
cord with long-range goals.
When they achieve a techno-
logical advantage in one area,
they use that advantage to in-
sure their advance into new
areas. They have targeted su-
percomputers as the next
high-tech area in which to es-
tablish a dominant position."
You couldn't tell it from
looking at the Japanese super-
computers in the U.S.,
though. The only one is an
NEC sx 2 leased by the Hous-
ton Area Research Consorti-
um (NARc). There have been
many other efforts to land a
Japanese supercomputer, but
none has succeeded (see "Su-
percomputer Dumping Al-
leged at U.S. Universities,"
Sept. 15, 1987, p. 17).
The SDSC would just as
soon keep it that way, but it
sees the U.S. government
tripping all over itself.
"The supercomputer
agreement with Japan con-
tains seeds which can further
undermine the position of
U.S. manufacturers," the re-
port contends. "It requires of
Japan a number of actions
which, if made reciprocal to
DATAMATION 0 AUGUST 15,1988 19
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
the U.S., could facilitate the
entry of Japanese supercom-
puters into the U.S. market."
Supercomputer Partnerships
Now, the supercom-
puter trade agreement re-
quires the Japanese in gov-
ernment or university super-
computer procurements to
give equal preference to U.S.
manufactured supercom-
puters. The Japanese essen-
tially have agreed to avoid un-
fair pricing by their manufac-
turers. According to
McAdams, however, what
looks unfair to a trade admin-
istrator looks to a manufactur-
er like recognition of partner-
ship with a government agen-
cy or university.
"This sounds like a com-
plete misunderstanding of the
trade agreement," says
Lauren Kelley, a supercom-
puter analyst in the Depart-
ment of Commerce (DOC)
Office of Computer and Busi-
ness Equipment. "The agree-
ment is based on the interna-
tional GATT [General Agree-
ment on Tariffs and Trade]
government procurement
code. In fact, nothing in the
agreement is different from
the standard General Ser-
vices Administration proce-
dures. To say this is a one-sid-
ed arrangement is completely
untrue."
Nonetheless, the IEEE
thinks a hard rain's gonna fall.
Here's the U.S. telling the
Japanese to open their mar-
kets or else, while simulta-
neously blocking the Massa-
chusetts Institute of Technol-
ogy (MIT) from purchasing an
sx 2 from NEC. Of course,
even a Freedom of Informa-
tion Act search wouldn't un-
cover a written policy on the
subject, but you can bet the
national debt that govern-
ment agencies aren't about to
open their doors to the
Japanese.
(The Department of De-
fense, which sees national se-
curity in every byte, is legisla-
tively prohibited from buying
Peak Supercomputer Performance Rates
SINGLE CPU PEAK ALLOWING MULTIPLE
64-BIT MFLOP RATE CPUS
Cray-1
160 160
Cray X-MP
Cray-2
Gay-3 (1989)
Cyber 205
ETA 10 (1986)
ETA 10/E
ETA 10/6(1988)
Fujitsu VP 100
Fujitsu VP 200
233 932(4 cpus)
488 1,952(4 cpus)
1,000 16,000 (16 cpus) est.
200(2 pipe) 400(4 pipe)
350 1,400(4 cpus)
415 1,660 (4 cpus)
643 5,142(8 cpus)
271 271
533 533
Fujitsu VP 400
1,067 1,067
Hitachi 5-810/20
630 630
titadii 5-820/80
2,000 N/A
IBM 3090/ VF
NEC SX 1
116 696 (6 cpus)
570 570
NEC SX 2
1,300 1,300
NK SX 3 (1989)
5,000 20,000(4 cpus)
Sourest Supercomputer Vulnerability,"
Subcomnittee.
any foreign?Congress
meant Japanese?supercom-
puters in 1988.)
"How long can such a
pattern last?" the paper asks.
"Is it realistic to believe that
the Japanese machines are
foreclosed from U.S. institu-
tions? If the requirements we
now impose on the Japanese
were made reciprocal for U.S.
universities and government
laboratories, Japanese super-
computers would have to be
acceptable to those agencies
on nondiscriminatory terms."
For some supercom-
puter users, that day can't
dawn soon enough.
"We Wont the Best Product"
"The economic leverage
of supercomputers is irrele-
vant and always will be,"
NsF's Karin says. "What mat-
ters is the use of supercom-
puters. We need the best su-
percomputers, and we need
to make the best use of them.
Who makes a supercomputer
is far less important than how
10I's Scientific Supercomputer
it's used."
"There's a genuine con-
cern here about foreign com-
petition," says a user at a ma-
jor federal lab. "But when it
comes to computers, we just
want the best product. And by
keeping out the Japanese, the
government and the U.S. su-
percomputer industry are
pretending the situation is
better than it really is."
However, by letting in
the Japanese, the IEEE sees
the industry living on desola-
tion row.
"The new requirements
could greatly facilitate the en-
try of the Japanese into the
U.S. At the same time such re-
quirements could disrupt the
implicit partnership between
U.S. manufacturers and U.S.
government laboratories
and/or U.S. universities, and
transfer the benefits of part-
nership to Japanese firms. In
response to a low bid, U.S. na-
tional labs could be required
to become partners to Japa-
nese firms in perfecting their
systems for penetration of
U.S. markets."
No U.S. lab would want
to do that, at least on the rec-
ord. But the Japanese clearly
have the fast single-processor
system and are expected to
increase that lead with their
next generation product ex-
pected next year (see "Peak
Supercomputer Performance
Rates"). So how much longer
can users be shut down at
their expense? Not very. So
it's only a matter of time be-
fore HARC has company.
Japan's Software Is Lacking
That could be very soon
if the Japanese get their soft-
ware act together. Their soft-
ware isn't quite up to their
hardware, but therein lies the
danger.
"Once general portabil-
ity of applications is achieved,
the avenue to continued mar-
ket leadership by U.S. firms
would be solely through tech-
nical leadership," the report
says. "But how can U.S. firms
simultaneously rely on coin-
ponentry manufactured by
their competitors, the Japa-
nese, and assure their cus-
tomers that they, the U.S.
firms, can maintain techno-
logical leadership?"
They can't.
"As soon as Japanese
firms have software that U.S.
companies need, they'll be
selling heavily here," the gov-
ernment official says. "NEC is
working its butt off to develop
software. When those devel-
opments take place, we have
no laws to restrict them."
This is generally expected to
be sooner rather than later.
So why not let them
come and fight it out nanosec-
ond-to-nanosecond in the
tried-and-true capitalist tra-
dition?
"Because our entire
economy is at risk," Mc-
Adams contends. "Super-
computers are the key to in-
dustrial design. If you lose su-
percomputers, you're in real
trouble."
20 DATAMATION 0 AUGUST 15,1988
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
SOFTWARE FOR SUPERCOMPUTERS
? A REPORT
Prepared by the Scientific Supercomputer Subcommittee
of the
Committee on Communications and Information Policy
United States Activities Board
INSTITUTE OF ELECTRICAL AND ELEC11ONICS ENGINEERS
1988
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
SOFTWARE FOR SUPERCOMPUTERS
A REPORT
EXECUTIVE SUMMARY
This is a summary of a report prepared by the Scientific Supercomputer Subcommittee of the
Committee on Communications and Information Policy, United States Activities Board, Institute
of Electrical and Electronics Engineers.
Early supercomputers
had poor software.
Portability, optimization,
algorithms.
Portability of code was
and is an important
problem.
The first truly unique supercomputer architectures started to appear
in the early 1970s. The Burroughs Illiac IV, the Texas Instruments ASC,
and CDC's Star 100 were built in small quantities and the software was
not highly developed. In many cases the purchasers provided most of the
software themselves. It is important to understand that the very unique-
ness that allows these computers to yield very high performance also
forces users to expend significantly more effort in optimizing their codes
to achieve even a fraction of this potential power. These early machines
were very difficult to use and the software was not easily optimized.
When the current class of supercomputers (called Class VI machines)
started to make their appearance in the late 1970s and early 1980s, new
software had to be provided. Because of difficulties in formulating and
optimizing the programs appropriately for these newer vector computer
architectures, the efficiency of use of Class VI architectures suffered (and
still does).
We have three main needs associated with supercomputer software:
portability; language and compiler related software, especially automatic
optimization; and architecture-appropriate algorithms.
The ability to take programs from one manufacturer's machines to
another or even to move code to later generations of the same equipment
is known as portability. This has been a continuing problem in the com-
puting industry, especially in supercomputing. Lack of portability not
only causes premature obsolescence of users' codes, but also shortens the
lifetime of system code supplied by the manufacturer, thus making it even
more difficult to justify the heavy cost of system code development.
Some early government users of supercomputers have tried over the years
to provide continuity from one generation to the next. For example, a
group at the Lawrence Livermore National Laboratory designed the Liver-
more Time Sharing System (LTSS) operating system in 1963 for the
-1-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Supercomputer software
is expensive, difficult.
Newer (multiprocessor)
architectures not only
make the problem worse
they create whole new
problems.
Better parallel algo-
rithms are needed.
Control Data Corporation 1604 computer and continued its development
through later CDC computers, including the CDC 3600, 6600 and 7600,
thus providing compatibility from generation to generation. The version
that runs on Cray computers is now called CTSS and is being used at
Livermore and by other DOE laboratories. This is one example of how a
national software center could help to provide some leadership in the area
of portability.
Writing good supercomputer software is especially difficult because
of the need to take advantage of the complex architectures needed for high
performance. For example, program optimization is an area of crucial
importance to supercomputers. Good automatic optimization is required
to achieve a higher percentage of the potential speed of supercomputers,
better utilization of scarce manpower, and better portability. A good deal
of work is now being done on automatic program optimization. But even
after using the best of these optimizers, the performance typically
obtained from the machine is far less than the peak performance possible.
Not only are vector optimizers not sophisticated enough, but simply
applying vectorization techniques to make use of multiprocessors is not
enough. Whole new techniques are needed to get acceptable multiproces-
sor performance.
Machines with new architectures possessing highly parallel struc-
tures are now being designed and built. At the moment, we are exploring
, the capabilities of high performance systems containing only a few paral-
lel processors. A number of supercomputer systems being planned are
somewhat larger, having up to 16 processors. Yet good optimization
software does not yet exist even for these low levels of parallelism.
Machines with new architectures possessing highly parallel structures
including hundreds, even thousands of processors, are now being designed
and built. Optimization for these machines promises to be even more
difficult and labor intensive than the last generation of machines. This
optimization is not just harder, it poses new problems not encountered
before. Efforts to design automatic optimization software to alleviate this
problem are at a very early stage and the costs involved in developing this
software are so high and the efforts to develop it are so fragmented, that
very little may ever see the light of day.
Better algorithms can make a major difference in the feasibility of
some applications. One only has to think of Fast Fourier Transforms
(FF 1) and the Simplex method to recognize the impact better algorithms
can have. Algorithms are especially important on supercomputers
because they need to be specially designed to take advantage of the vector
and multiprocessor parallelism. Fortunately, some more economical com-
puters now available make it possible to experiment with new parallel
- 2 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Need for coordination.
algorithms, and many of these systems are being used in this fashion.
However, we are a long way from saying that we know how to use paral-
lel processors efficiently for most problems. We now have the tools to
study parallel algorithms and we must make these tools available to the
algorithms community.
Many of our difficulties stem from a lack of coordination of effort.
Manufacturers were reluctant to cooperate out of fear of antitrust laws,
and they were reluctant to finance significant software development for
what they viewed as a short term product. When the Japanese entered the
supercomputer competition, they took a more global approach to super-
computers and treated them as important, highly marketable entities.
Having established supercomputers as a national priority, they were able
to take a longer term view of software development for these machines, a
policy which is proving to be successful. The performance of Japanese
machines shows the results of their foresight by outperforming the U.S.
supercomputers in many instances (see Table 2), despite their relatively
recent entry into the supercomputer arena. One should wonder how the
performance of U.S. machines will compare with our competition in a few
years.
Recommendations.
Risk sharing, coordina- This subcommittee makes five recommendations aimed at improving
lion, and national the health of the U.S. supercomputer industry. These recommendations
laboratories, include more participation from the federal government in the financial
risks involved with supercomputer hardware and software development,
better coordination of U.S. activities in this area, and the establishment of
national software laboratories whose purpose would be to provide
software specific to supercomputing.
-3-
-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
SOFTWARE FOR SUPERCOMPUTERS
This is a report prepared by the Scientific Supercomputer Subcommittee of the Committee on
Communications and Information Policy, United States Activities Board, Institute of Electrical
and Electronics Engineers.
IEEE.
Background.
Need for strong govern-
ment role.
1. Introduction.
The Institute of Electrical and Electronics Engineers (IEEE) is the
world's largest engineering society, with over 295,000 members world-
wide, 223,000 of whom live and work in the United States. The United
States Activities Board of the IEEE takes the position that advanced
scientific computing capability is a technology that must be accelerated in
the United States. This technology is crucial for national defense,
economic growth, and advances in engineering and science.
In 1983, the United States Activities Board of the IEEE set up a spe-
cial committee to examine the U.S. position in supercomputer develop-
ment and to recommend actions the government should take to ensure
continued preeminence of the U.S. in this challenging and complex field.
Among its five recommendations, the committee proposed that the Federal
government make a long-range commitment to maintaining leadership in
supercomputer development and take an active role in fostering develop-
ment of new systems.' The establishment of this committee on supercom-
puters and the continuing support of the its recommendations demonstrate
the continuing deep concern of the IEEE on this matter.
This subcommittee reiterates the position taken in the prior report
and restates, in the strongest terms, its own belief that the government
should pursue an active role. Supercomputer software is not just impor-
tant to the position of the U.S. supercomputer industry in the world
market, but it is also crucial to a much broader spectrum of industries that
depend on supercomputers for the design of competitive products, not to
mention its strategic value to U.S. defense. Toward this end, the govern-
ment must allocate sufficient research funds to improve components,
architecture, systems software, applications software, and very high per-
formance peripheral equipment.
I "Scientific Supercomputer Committee Report," produced by the Scientific Supercomputer Subcommittee of the
IEEE U. S. Activities Board, Sidney Fernbach, Chairman, October, 1983.
- 4 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Slow progress since
1983.
Early machines evolu-
tionary, good software
evolved from existing
products.
Limited-production
machines had poor
software.
Exception: 6600 suc-
cessful, users helped
with the software.
Newer architectures
require more effort to
optimize.
In the intervening years since October 1983, much progress had been
made on many fronts. New programs budgeted in the tens of millions of
dollars have been started in several agencies including DOE, NSF, NSA,
and NRL. Prior efforts have been redoubled in all three arenas: the
universities, the private sector, and the Federal government. These are
hopeful auguries, but the advances are only incremental. Neither these
steps nor those currently being considered will be enough. Thus, this sub-
committee further proposes a particular action agenda to enable the
universities, the private sector, and the government to work together to
insure our preeminence in supercomputers.
2. Historical Background.
The definition of a supercomputer is that it is the most powerful
scientific system available at any given time.2 Early machines in this
category, e.g., the IBM 7094, CDC 6600, and IBM 360/91, topped off the
standard line of the manufacturer. Hence, based upon the fairly large
sales volume for the low end of the equipment, a reasonable amount and
quality of software was supplied with them.
With the advent of "one-of-a-kind" machines, the situation started to
change somewhat. The IBM NORC (only one existed), the Sperry-Univac
LARC (only two were built) and the IBM STRETCH (nine were built)
each had unique software that was incompatible with that of other sys-
tems. These machines tended to be revolutionary rather than evolution-
ary, and they lacked the market volume to permit the extensive software
development required for a revolutionary machine.
An exception to the rule, and perhaps the first highly successful
supercomputer was the CDC 6600. Control Data Corporation (CDC), in
an important symbiotic relationship with some of its customers (the
national laboratories), was able to provide software for the system and
went on to sell a large number of machines. The successor to the 6600
was the CDC 7600. The compiler for the 6600 formed the basis for the
7600 compiler, and with this head start and more help from a few custo-
mers, adequate software was again available. Thus many copies of the
CDC 7600 were sold, and it too became quite successful.
Then even newer architectures began to emerge, architectures which
required far greater effort to restructure or optimize programs to utilize the
higher power available in these machines. When the current class of
supercomputers (called Class VI machines) started to make their appear-
ance in the late 1970s and early 1980s, the software had to be redone yet
2 Supercomputing ?An Informal Glossary of Terms. IEEE, 1111 Nineteenth Street, NW, Washington, DC, 1987.
-5-
-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Why manufacturers have
not provided good
software for supercom-
puters.
Broadening market and
foreign competition
require better software.
Why the Japanese are
successful.
again. These machines were very difficult to use and the software was not
easily optimized. In many cases the purchasers provided most of the
software themselves. Because of the difficulties of restructuring problems
to take advantage of the vector oriented machine architecture, much of the
potential of these machines failed to be realized. It is not unusual for sus-
tained performance on these machines to be less that 20% of the peak per-
formance. Once again we see that industry has not had the immediate
market base to justify this software development, nor the lead time to
accomplish this development.
3. What do we need?
Because of their cost and complexity, the market for supercomputers
has been small and, additionally, writing systems software for these
machines has been particularly difficult. The result has been a
significantly lower standard for scientific supercomputer software, a stan-
dard we have been able accept only because of a small and sophisticated
cadre of users. For years manufacturers were able to deliver their pro-
ducts to the few customers who could afford them--customers who could
also afford to do most of the software development themselves. Because
of the limited volume of the market, short viewpoints of the financial
community, and limited lifetime of the software (due to lack of portabil-
ity), expensive software development projects could not be justified.
Industry is now beginning to realize that there are important
economic advantages in exploiting modern supercomputers. Many new
supercomputer installations are being set up to satisfy the needs. See
Table 1 for current or currently planned installations. Now that the
market for supercomputers is broadening, new classes of users require a
much higher standard, especially if our industries are to benefit from the
superior design capabilities afforded by supercomputers.
When the Japanese started to build supercomputers, they recognized
this broadening market and they took a more global approach to super-
computers, treating them as another important, marketable commodity.
Having established supercomputers as a national priority, they were able
to take a longer term view of software development for these machines, a
policy which is proving to be successful. (See Table 2 for a comparison
of performance of certain computers as on the LINPACK programs as
evaluated by Jack Dongarra of Argonne National Laboratory.) Obviously,
the Japanese manufacturers have recognized the areas in which they
should apply strong efforts. One should wonder how the performance of
U.S. machines will compare with that of our competition in a few years.
-6-
-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
TABLE 1
FREE WORLD DISTRIBUTION
OF SUPERCOMPUTERS
(INSTALLED OR ON ORDER AS OF 12/31/87)
COUNTRY
NUMBER OF SUPERCOMPUTERS
UNITED STATES
134
JAPAN
70
UNITED KINGDOM
19
GERMANY
17
FRANCE
16
CANADA
7
HOLLAND
3
NORWAY
3
SWEDEN
2
SWITZERLAND
2
ABU DHABI
1
AUSTRALIA
1
ITALY
1
SAUDI ARABIA
1
TAIWAN
1
FREE WORLD DISTRIBUTION
BY APPLICATION
(APPROXIMATE)
APPLICATION:
NUMBER OF SUPERCOMPUTERS
Research:
81
Defense:
45
Universities:
45
Aerospace:
32
Petroleum:
30
Weather:
16
Nuclear Energy (weapons, reactors):
12
Automotive:
'
11
Service Bureaus:
9
-7.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
TABLE 2
COMPUTER PERFORMANCE
SOLVING A SYSTEM OF LINEAR
EQUATIONS WITH LINPACKt
(FULL PRECISION--ALL FORTRAN)
SYSTEM
MFLOPS
*ETA 10-E (1 proc. 10.5ns)
52
*NEC SX-2
43
*CRAY X-MP-4 (1 proc. 8.5 ns)
39
*NEC SX-1
36
*NEC SX-1E
32
*CRAY X-MP-2 (1 proc.)
24
*CRAY-2 (1 proc.)
21
*AMDAHL 1200
19
*CDC CYBER 205 (2-pipe)
17
*FUJITSU VP-200
17
*HITACHI S-810/20
17
*CRAY is
12
*IBM 3090/180 VF (1 proc.)
12
FUJITSU M-380
6.3
CDC CYBER 875
4.8
AMDAHL 5860 HSFPF
3.9
CDC 7600
3.3
IBM 3090/120E
3.1
CONVEX C-1/XP
3.0
FPS-264/20 (M64/50)
3.0
IBM 3081 K (1 proc.)
2.1
HONEYWELL DPS 8/88
1.7
AMDAHL 470 V/8
1.6
IBM 370/168 (fast mult.)
1.2
AMDAHL 470 V/6
1.1
ELXSI 6420
1.5
DEC VAX 8600
.48
IBM PC (W/8087)
.012
*These machines are generally acknowledged to be supercomputers.
tData from J. Dongarra in "Performance of Various Computers Using Standard Linear Equations Software in a For-
tran Environment," Table 1 (February 2, 1988, Linpack, full precision, no Blas), Computer Architecture News, Vol.
16, No. 1, March, 1988.
- 8 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Supercomputer software
is difficult to write.
Newer (multiprocessor)
architectures make the
problem worse.
What do we need?
Parallel algorithms.
Writing good supercomputer software is especially difficult because
of the need to take advantage of the complex architectures needed for high
performance. For example, program optimization is an area of crucial
importance to supercomputers. Good automatic optimization is required
to achieve a higher percentage of the potential speed of supercomputers,
better utilization of scarce manpower, and better portability. A good deal
of work is now being done on automatic program optimization. But even
after using the best of these optimizers, the performance typically
obtained from the machine is far less than the peak performance possible.
Not only are vector optimizers not sophisticated enough, but simply
applying vectorization techniques to make use of multiprocessors is not
enough. Whole new techniques are needed to get acceptable multiproces-
sor performance.
Machines with new architectures possessing highly parallel struc-
tures are now being designed and built. At the moment, we are exploring
the capabilities of high performance systems containing only a few paral-
lel processors. A number of supercomputer systems being planned are
somewhat larger, having up to 16 processors. Yet good optimization
software does not yet exist even for these low levels of parallelism.
Machines with new architectures possessing highly parallel structures
including hundreds, even thousands of processors, are now being designed
and built. Optimization for these machines promises to be even more
difficult and labor intensive than the last generation of machines. This
optimization is not just harder, it poses new problems not encountered
before. Efforts to design automatic optimization software to alleviate this
problem are at a very early stage and the costs involved in developing this
software are so high and the efforts to develop it are so fragmented, that
very little may ever see the light of day.
What do we need? 1) Better portability, so that software has a longer
lifetime and can therefore sustain more development; 2) Better program
optimizers, so that users can spend their time more productively, and get a
higher fraction of the potential power of the supercomputer; 3) Better
algorithms, again to achieve a higher fraction of the potential power; 4)
Better languages and operating systems, to improve ease of use, expres-
sive efficiency, and execution efficiency;
Better algorithms can make a major difference in the feasibility of
some applications. One only has to think of Fast Fourier Transforms
(It 1) and the Simplex method to recognize the impact better algorithms
can have. Algorithms are especially important on supercomputers
because they need to be specially designed to take advantage of the vector
and multiprocessor parallelism. Fortunately, some more economical com-
puters now available make it possible to experiment with new parallel
-9-
-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
?
algorithms, and many of these systems are being used in this fashion.
However, we are a long way from saying that we know how to use paral-
lel processors efficiently for most problems. We now have the tools to
study parallel algorithms and we must make these tools available to the
algorithms community.
Better languages are Once algorithms have been designed, better languages must be pro-
needed. vided to allow more efficient expression of these algorithms and to allow
their efficient execution. Fortran is the traditional language of scientific
computation and a new standard incorporating vector extensions is
expected to be out soon. Two reasons for the popularity of Fortran, even
in the face of its age, are its execution efficiency and its portability.
Unfortunately its portability is limited: It is easy to move a Fortran pro-
gram from one machine to another, but the level of the machine dependent
detail which users put into their program to gain execution efficiency
often does not prove effective on another machine. Low level details
included in a program to improve execution efficiency on one machine
may (and often do) prove detrimental on another machine. Thus the pro-
gram may execute correctly on another machine, but with considerable
loss of efficiency. In any event, other languages and programming para-
digms are proving to be as portable as Fortran and considerably more
expressive.
Portability. One does not wish to program all problems for all machines, espe-
cially when it means reprogramming each program in order to get
optimum performance. But there has been little or no compatibility
between supercomputer systems. Originally, the manufacturers desired it
this way. It was easier to lock a customer into a given series of machines
by making conversion to another much too difficult. This may be good
for individual manufacturers, but certainly makes for great difficulty at the
user level, whether it be in government or industry. The resulting loss of
productivity is not good for either the customer or the U.S. economy. For-
tunately, driven by forces like the popularity of the Unix operating system
and the demands of users, manufacturers are changing their ways. Porta-
bility has a chance of becoming a reality. But portability requires much
more to be done both in the area of standards and optimization.
Advantages of portabil- Portability will greatly extend the lifetime of software, both applica-
ity. dons software and systems software. One only has to look at the growth
of Unix to see the potential for portable operating systems. Given these
longer lifetimes, we can afford considerably more effort to achieve better
software, and users will be able to spend more of their time on truly
creative work like designing new algorithms.
-10 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Optimization. True portability requires a high level representation of algorithms,
with no machine dependent semantics. However, a sophisticated program
optimizer is needed to get the required machine efficiency. Thus, for
example, details of memory hierarchy management?vector registers,
cache, virtual memory? should not be part of the user's program but
must be optimized by the software itself. This allows true portability, and
also permits more productive use of the programmers' time.
4. How do we get what we need?
Earlier reports and During the early 1980s it was realized that the U.S. scientific effort
actions, was suffering from a lack of adequate facilities to carry out large scale
research projects. The Lax Report3 to the National Science Board pointed
out the lack of availability of supercomputer resources for researchers in
our universities. Through supercomputer center grants from NSF, this
situation is beginning to be corrected. We are placing hardware at many
sites, so that availability is being increased significantly. The Lax Report
also urged the need for training in the use of supercomputers, and for
research and development to plan future generations of supercomputers.
Little has been done in these areas, as yet. A more recent report4
reiterates these concerns.
Standards. The IEEE and similar organizations should play a larger role in
establishing standards for languages and operating system software that
would improve portability, similar to the role played in the establishment
of the Posix standard for Unix. In the supercomputer marketplace, the
government's share of the market is close to 41%. It is expected that the
U.S. Government would do its part by adopting these standards and
enforcing them on government purchases. Such standards will enhance
portability, longer software lifetimes, and broader markets.
Incentives and disincen- In the past, most software was written by the hardware vendor or by
tives, the customers, but this is beginning to change. Software is increasingly
provided by third parties, independent of the vendor's particular hardware.
Given a larger, multivendor market for software, and the resulting
increase in motivation for portability, better software results. But the
development costs for .supercomputer software are still enormous, and
most software houses prefer to concentrate on markets like the personal
3 Report of the Panel on Large Scale Computing in Science and Engineering, Peter D. LAX, Chairman, National
Science Foundation, December, 1982.
4A National Computing Initiative: The Agenda for Leadership, produced under the auspices of the Federal Coordi-
nating Council for Science, Engineering, and Technology (FCCSET), sponsored by NSF and DOE, and published
by SIAM, Philadelphia, 1987.
-11-
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
National software
centers.
Importance of super-
computers and need for
government directed
focus.
computer market where the volume is high and the software easy. Incen-
tives from the government such as guaranteed purchases (software and
hardware) and even direct development contracts are needed to focus
some of the attention of the third party software vendors on supercom-
puter software. Additionally, vendors' investments in software need ade-
quate protection under the law. And finally, not only should any impedi-
ments to cooperation between vendors (hardware and software) and custo-
mers (especially national laboratories) be removed but such cooperation
should be encouraged. Stimulation of better software for supercomputers
and the protection of the investment in this software development involves
many complex issues beyond the scope of this report. These issues need
to be discussed and resolved.
Additionally, a mechanism needs to be set up to both better focus our
efforts in the area of supercomputer software and to take on the most
expensive and risky elements of software development, much as the
national laboratories take on these risks in the areas of energy, weather
research, and health.
5. Recommendations.
Supercomputer software is important not just to the position of the
U.S. supercomputer industry in the world market, but is also crucial to a
broad spectrum of industries that depend or will come to depend on super-
computers for the design of competitive products. This subcommittee
believes that the importance of supercomputers in government and indus-
try is just being recognized. Nevertheless, software for supercomputers
remains underdeveloped due to the relatively small size of the supercom-
puter software marketplace (compared, for example, to the market for
workstation and personal computer software) and the fragmented and
uncoordinated efforts in this area. Although some attempts have been
made to remedy the situation, we believe that it would be in the best
interests of the United States if the government were to provide more
focus on this problem through the following actions:
(1) Stimulate the supercomputer industry by underwriting some of the
costs and risks of hardware and (especially) software development.
This might be done through a program where the government, during
the early stages of the development cycle, commits to purchase
supercomputer systems (providing they meet certain performance
requirements). Not only would this help to underwrite risks associ-
ated with these machines, but would provide a stronger voice in their
design.
- 12 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
(2) Improve the state of supercomputer software by direct research and
development contracts and grants to industry and government labora-
tories.
(3) Increase basic research funding in supercomputer software.
(4) Establish a formal coordinating body to better focus existing
development efforts through standards for software portability, and
to provide interagency coordination of Federally funded research
efforts.
Establish several laboratories, such as the National Supercomputer
Software Research and Development Institutes recommended in the
SIAM report5 and in an earlier report from this subconirnittee,6 to be
associated with existing supercomputer centers, both academic and at
national laboratories. The early successes of some of the present
national laboratories is evidence of the potential for success of such
institutes. These institutes should have the following goals.
? Advise the Federal government on matters relating to
supercomputing;
? Set common software specifications for supercomputers;
? Carry out practical research in structuring algorithms and
applications for supercomputers, including parallel (mul-
tiprocessor) algorithms;
? Develop software packages, including operating systems
and compilers, that would be suited for a wide variety of
supercomputers;
? Devise performance measures for supercomputers; and
? Package these products for U.S. government, educational,
and industrial use.
(5)
5 Ibid.
6Software for High Performance Computers, Prepared by the Subcommittee on Supercomputers of the Committee
on Communications and Information Policy of the Institute of Electrical and Electronics Engineers, December,
1985, Washington D.C.
- 13 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Participating Members of the Subcommittee on Scientific Supercomputing
Sidney Fernbach, Chairman
Heidi F. James, Committee Secretary
Alfred E. Brenner
James F. Decker
Duncan H. Lawrie
Alan McAdams
Kenneth W. Neves
John P. Riganati
Stewart Saphier
Paul B. Schneck
Lloyd Thorndyke
Hugh Walsh
The IEEE Subcommittee on Supercomputers of the Committee on Communications and Infor-
mation Policy has produced a number of position papers on supercomputers. For further infor-
mation, or to be placed on a continuing mailing list, contact:
Heidi F. James
IEEE Washington Office
1111 19th Street, N.W.
Suite 608
Washington, D.C. 20036
(202) 785-0017
Production: ? Fri. May 20 17:29:06 CDT 1988
File: uicsrdsw:IxdO0flhomesllawriellEEE-CSIFernbachlsoftware.position
- 14 -
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
1
Science Revision: 5/16/88
THE COMPUTER SPECTRUM:
A PERSPECTIVE ON THE EVOLUTION OF COMPUTING
IEEE Scientific Supercomputer Subcommittee
Since the mid 50's, the modern computer has rapidly evolved to satisfy the computational
needs of an increasingly large fraction of society. This evolution is discussed and the
development of the various classes of computers that have come into being to satisfy the
,increasingly diverse computational needs which have developed is explored. Based upon
this historical analysis, projections for the near future directions of growth are offered.
Computers in widespread use today span a range of sizes and a range of applications
much greater than that of any other manufactured product in our highly technological
society. Quantitatively, the costs of manufactured items which might properly be called
computers range from $20 for hand held programmable devices to more than $20 million
for the largest supercomputer systems in the marketplace today - a factor exceeding 106
in cost and a similar factor in terms of the computational power and storage capacity.
Although their range is not quantifiable, the breadth of applications for which these
computers are put to work is equally impressive.
The IEEE Scientific Supercomputer Subcommittee is a technology-policy subcommittee
of the IEEE United States Activities Board, 1111 19th Street, N.W., Washington, D.C.
20036. The subcommittee members participating in the development of this paper are:
Alfred E. Brenner, Supercomputing Research Center, Lanham, MD
Sidney Fernbach (Chairman), Consultant, San Jose, CA
Duncan Lawrie, Center for Supercomputing R&D, University of Illinois, Urbana, IL
Alan McAdams, Cornell University, Ithaca, NY
Kenneth W. Neves, Boeing Computer Services, Seattle, WA
John M. Richardson, National Research Council, Washington, D.C.
John P. Riganati, Supercomputing Research Center, Lanham, MD
Stewart Saphier, Department of Defense, Washington, D.C.
Paul B. Schneck, Supercomputing Research Center, Lanham, MD
Lloyd Thorndike, ETA, St. Paul, MN
Hugh Walsh, IBM Data Systems Division, Kingston, NY
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
2
To satisfy this broad spectrum of use and to accommodate to the pocketbooks of the
broad range of potential customers, manufacturers of computers continue to develop at a
dizzying rate new types of computers and specialized software to satisfy particular
market niches. In fact, the options available are an elegant example of the diversity that
mass production has produced in the modern world as so pointedly observed by Alan
Toeffier in his book Future Shock (1). Although there is a multitude of terms in use
today identifying classes of computers, there seems to be little precision or even
? agreement on these classifications.
This paper offers a perspective on appropriate classifications for computer systems and
gives a historical review of their evolution to the present time. Then, with the hindsight
of this historical analysis, projections into the near future are made giving what further
changes and evolutions may be expected.
The First Two Computer Classes: Business and Scientific
In the early days of the modern computer era, i.e., during the decade of the 1950's, all
machines were thought to be equally useful on all problems. It should be noted,
however, that while the range of applications was quite diverse, the extent of use at that
time was quite limited. The computers used either binary or decimal numerical
representations. Soon, two marketplaces developed. One for business applications and
the other for scientific applications. In the former case, fixed-point decimal arithmetic
seemed in order; for the latter, floating-point binary arithmetic seemed preferable. These
differences persist, more or less, until the present day.
During the latter part of the 50's and early into the 60's, the two marketplaces were
developed more or less as separate entities with separate cultures. Typically computer
architectures focused on one or the other of the two marketplaces. However, starting
with the announcement (2) by IBM of the System 360 in 1964, there was an attempt by
manufacturers to coalesce at least the hardware (3) satisfying these two marketplaces (4).
Thus the architecture of the medium to large-scale computers developed in the mid-60's
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
3
included both decimal and floating-point arithmetic. Although unified operating systems
were introduced to satisfy both communities, separate high level languages and
applications packages were developed to satisfy the specific needs of each of these two
major marketplaces. These machines continued to evolve through the 70's and 80's into
quite general purpose computing engines, approximately equally capable of serving both
classes of use. Today these have come to be called mainframe computers.
The Emergence of New Computer Classes; the Supercomputer and the Minicomputer
Starting in the 60's, there was a persistent and major demand for scientific computers yet
more powerful than the mainframes then extant. This new market niche, driven by the
needs of the U.S. Government for national security purposes and by the need of the
meteorological and seismic communities, was developed at the very high end of the
scientific computing spectrum. The computers serving this market niche, the most
powerful scientific computational systems available, have come to be called
"supercomputers" (5).
The demand for ever faster computer performance for scientific problems has persisted
for over four decades. Approximately in the mid 1970's, the capabilities of
supercomputers crossed a critical threshold allowing "computational science" to develop
explosively. Computational science now stands, de facto, alongside the theoretical and
experimental sciences as a fully legitimate field. In almost every scientific and
engineering discipline, technology development is coming to rely more and more heavily
on computer simulation, since only a small fraction of desirable experiments can be
carried out physically in a cost-effective or completely realistic way. This has fueled the
ever-increasing demand for more scientific computing power.
Early in the modern computer era, the technical and business communities quickly rose
to the challenge of putting computers to active use. However, the early systems were all
physically quite large and very costly. High cost inhibited the proliferation of these
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
4
(mainframe) systems. Yet, a very large potential marketplace awaited computers with
price tags substantially below those of the then-available mainframes. A new class of
computer system, the forerunners of today's minicomputers, with lower computational
speed but more than proportionally reduced cost were introduced, primarily by new
companies. Early examples of these were Digital Equipment Corporation's PDP 1 (6)
and IBM's 650 (7). Other entries from additional manufacturers entering the field soon
followed. Thus, by the mid-60's, although the general scientific and business
marketplaces were still -distinct, there were now three products available in the
marketplace: minicomputers, mainframes, and supercomputers.
As with all new products, variations soon began to appear to meet the requirements of
different niches in the marketplace. The major categories remained, but new systems
were introduced with variations in performance, size and kind of memory, word length
and types of instructions. Software was further diversified in an attempt to effectively
address the particular needs of the wide variety of users in the marketplace. Price, size,
ease of use, and relative performance remained the major factors determining the success
of the various offerings. The price was the most important single factor, a feature that
persists to this day. The minicomputers, which established a price niche, have persisted
to this day with price tags which span a very wide range, from quite small to the high end
superminicomputers which are larger and more capable than many of the current low-end
mainframe systems.
The Evolution of the Supercomputer
In the 1970's, two new architectural features - vector processing and parallel processing -
that would have widespread repercussions made their first appearances. These were
typified by the STAR 100 (9) - a vector machine (10) built by Control Data Corporation,
and the Illiac IV (11) - a parallel processor (12) designed at the University of Illinois and
built by Burroughs Corp. Vector processing was the first of the two to be introduced into
commercially available supercomputers. It gave rise to major performance
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
5
improvements. However, additional gains achievable through architectural,
technological, and software improvements to this approach are reaching saturation. It is
now quite apparent that major additional increases in computational power using current
technology must be derived primarily from parallel processing approaches. All high
performance computer vendors have begun to introduce this approach into their product
lines.
? Over the years, continuing advances in technology have allowed for increasing
supercomputer performance while total system prices have remained essentially constant.
Typically, supercomputer acquisition costs are in the range of from 10 to 20 million
dollars. Despite the rapidly falling price/performance ratio, the high acquisition cost
continues to be a factor inhibiting more widespread use of such systems, even in those
cases where the life cycle payoff may far exceed the investment.
In an attempt to make supercomputers more readily available to a broader group of users,
high performance, wide-area networking is now being introduced to allow for effective
remote access to these costly facilities. Communication lines operating at the highest
speed commercially available interconnect user equipment at local sites to remotely
located supercomputers. Most often, the communications are handled by front-end
computers or other smaller computer systems at the supercomputer site so as not to
burden the supercomputer with communications overhead. The computers used as front-
ends span a wide spectrum of capabilities, depending on what is readily available or
affordable to the given user.
Because of the high cost of supercomputers and/or the additional complications involved
in accessing such facilities remotely, there are continuing strong pressures to find
alternatives to solving problems by means other than supercomputers. This gives rise to
a continuous search for other more readily available, typically less expensive, computing
resources. When smaller mainframes or even minicomputers are adequate to the
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
6
problem solution, such alternatives do work well, but only because these problems do not
require supercomputers. Nevertheless, there remain large (growing) numbers of
problems which are sufficiently demanding in their computing requirements that only
supercomputer performance level facilities are appropriate.
The Emergence of the Minisupercomputer
By the 1980's, these pressures gave rise to yet another category of computer, the
"minisupercomputer." These are relatively inexpensive, readily available, high
performance computing engines which are architecturally similar to supercomputers, but
typically use the manufacturing methods of minicomputer vendors. Their cost range
makes them available to small groups of users (departments) within a company or
university. Great incentives existed for vendors to satisfy this newly developing market
niche. As has happened in the past, new players perceived the developing market niche
and quite a large number of new firms are trying to fill the gap.
The minisupercomputer is a high performance computing engine with a much lower cost
than that of a leading edge supercomputer. In many senses, this gives rise to the
"personal" supercomputer. The engineering trade-off here is in favor of economy - not
performance - in contrast to the performance over cost trade-off which drives
supercomputer design. There is a very high demand from the research community and
industry for such economical machines. It is too early to tell how broad is the funding
base to support this class of need. This approach appears to be the "poor" (not wealthy)
scientist's and engineer's preferred alternative to the remote access to a large
supercomputer. A major factor in this reference is the advantage of local rather than
remote control of the resource.
Two technological factors are having a major impact on the continuing development of
the minisupercomputer. These are the microprocessor and parallel processing. The
microprocessor - a full fledged processor on a chip - made possible the development of
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
7
the low-cost personal computer whose sales now number in the millions of units.
Revenue from these, in turn, have fueled the development of even more powerful
microprocessors. Secondly, the advancing technology utilized by successive generations
of supercomputers has slowly been reaching the natural limits imposed by the speed of
light and material limitations for the processees which have become the basis for
increasing technological improvements. Consequently, major future improvements must
come from an approach which attains high performance by architectural designs capable
of parallel processing on a single computation. This is the direction taken by all
manufacturers involved in high performance computing, and includes large mainframes
as well as supercomputers. Also, the juxtaposition of microprocessors into parallel
processing architectures have made it possible for many universities and many small
firms to engage in low-cost experimentation on parallel processing. This has given rise
to a large number of quite modest, mostly start-up, commercial entities competing for a
broadening minisupercomputer niche.
The minisupercomputer will not supplant the supercomputer. There will always be a
need for larger and faster machines. However, there will also be a need for less costly
systems (minisupercomputers) to serve as powerful "front-ends," stand alone systems,
and distributed departmental level systems.
Other Important Classes; The Personal Computer and the Applications Workstation
Although not the first entry in the arena, Apple Computer introduced its inexpensive
user-friendly personal computer (13) in 1977. It rapidly established a new niche in the
marketplace. This resulted from the low cost and quite powerful processing capability of
the product. Most importantly, the new systems required very little prior education on
the part of the user for their successful application to problems. IBM, in 1981,
recognizing the major new market potential, also entered this mass market low-cost
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
8
product arena. These two firms are now the major players in this niche, a niche quite
removed from all previously identifiable classes of computers.
The personal computer started as an expensive toy. Electronic games were a major use in
their early days. With time, more serious applications activities, especially word
processing, spreadsheets, and data base procedures, quickly became the primary function
for these machines. Now personal computers can be found in many homes. They have
been the basis for the transformation of the office into today's "electronic office." The
breadth of applications to which these machines have been put has increased enormously.
Theis range of performance and price now also spans a wide range.
Even more recently, specialized, very powerful classes of personal computers have
entered the marketplace. Initially these were "scientific workstations," but more
generally "applications workstations." These machines span computing performance
levels from the high end of the personal computer well into the arena heretofore covered
by the minicomputer and low end mainframes. The driving force for the development of
this new class of system is primarily the rapid growth in the performance capability of
mass produced microprocessors. Powerful processors on a single silicon chip that are
mass produced at remarkably low cost have made it possible to place significant
computing capability into a desk-top package. Only a few years ago, systems of similar
power required investments in excess of $100,000 and took up a good fraction of a room.
Scientific and engineering workstations have blossomed into a major new industry over
the last few years. This niche, too, is now well established. Low cost, high performance,
personalization of such workstations to an individual's needs, and their dedicated
availability to an individual make them a very attractive product.
Historical Analysis
In this section, the evolution of the computer industry is traced through the stages of its
development. The bases for this representation of the industry are: the applications
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
9
served, the cost of an individual system and the total annual revenues achieved in each
market segment. Earlier, particular niches were identified as they developed, and the
factors that influenced their development, growth and evolution were discussed. Figure 1
shows snapshots of computer industry development representative of each period.
In these graphs, the "applications" axis is a qualitative one, divided into "scientific" and
"business" applications. The "unit cost" axis (14) is the price of the various systems
stated in thousands of dollars of the year of the given snapshot. The vertical axis is the
"total annual revenues" for each of the market segments depicted, again stated in dollars
of the year of the given snapshot. The latter two axes use logarithmic scales to
accommodate the range.
Figure 1(a) shows the situation as of 1960. The dollar volume of computer revenue that
year was almost $700,000,000. At that time, there was a well-defined bifurcation of
scientific and data processing (or commercial) computing, each served by different
computer models.
In the ten year period between 1965 and 1975, the minicomputer market segment
emerged and began to take on a substantial fraction of the workload. Also during that
period, as shown in Figure 1(b) for 1970, the emergence of the general purpose
mainframe allowed for the coalescence of the scientific and commercial marketplaces.
During the period 1975 to 1980, with 1977 shown in Figure 1(c) as a representative year,
the supercomputer market segment emerged. It responded to a need at the high end of
the scientific computing applications space. Also shown is the beginning of the personal
computer market segment, quite separated in its early days from the rest of the "serious"
computing systems.
During the first half of the eighties, there was continued growth in all market segments;
Figure 1(d) depicts the situation for 1982. The scientific workstation emerged spanning
6. Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
10
the space between the personal computer and the minicomputer. Finally, the current state
of affairs is shown in Figure 1(e); 1987 is shown, although the data here is preliminary.
There is continued growth in all computing market segments and the development of the
latest niche -- the minisupercomputer -- has started.
It should be noted that at every step along the way, there have been other . classes of
computers that were developed, but failed to achieve significance as market identifiable
segments. Some of these, e.g., attached array processors, are still being marketed.
Because their dollar volume has remained limited, however, they do not appear in the
graphs. In many cases, uses of array processors have gradually been satisfied by other
products.
The current annual revenues of the U.S. computing industry are in excess of 100 billion
dollars. The industry continues as one of the fastest growing components of the U.S.
economy. Given the entrepreneurial spirit, especially strong in the electronics and
computer components of our industrial society, there continues to be new attempts to
develop heretofore unrecognized niches. Many of these entrepreneurial efforts fail either
because the ideas are unsound or badly implemented, or as frequently happens the timing
is not right. Nevertheless, the activity can be expected to continue, evolving in the
marketplace, and establishing new niches.
The Missing Element in the Analysis: Software
Making effective use of any hardware system requires appropriate software. In this
arena, the level of maturity is well behind that of the hardware. To some extent, this is
related to the lower level of maturity of our software experience relative to our hardware
experience. The construction of electronic hardware items in our industrial base goes
back over half a century. Our schools address hardware in a mature engineering and
production sense in the light of this experience. This is not the case with software.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
11
One of the problems is that not only is software much newer in concept, but also the
options available in its implementation are much greater than those in the hardware
arena. As contrasted to the situation twenty or more years ago when the early computer
systems were delivered, the software efforts required to build a system today are much
more demanding than are the efforts for the hardware development. Most firms have not
adequately integrated this new relationship into their planning and development cycles.
It is the software more than the hardware that the user sees in interactions with computer
systems. Sensible developments of software are further hampered by the fact that there
are few well agreed upon standards to which the designers must conform. In many cases,
to differentiate their products, manufacturers will resist agreeing to proposed industry-
wide standards. Some will explicitly avoid using a standard even when it does exist.
These are serious problems which are hampering the growth of the computing industry.
It is important to facilitate the maturation of the software discipline to bring it up to the
level of the hardware disciplines in use today. Until this is accomplished, investments in
computing will not give rise to optimum return. Since supercomputers play an important
role in advancing U.S. technology, it is imperative that these problems be solved.
Unfortunately, the short-term need of profitability on the part of industrial providers is
not always consistent with fostering these developments. Thus, because of the
importance of software, there is a need for government to take an active part in fostering
research, development, and industrial coordination to bring this about (16).
The Future Spectrum
What projections can be made for the evolution of the computer industry during the next
decade? One of the important developing arenas is that of universally available
networking. Computer networks are becoming important in every facet of research and
industry. The merging of the communications and the computing industries continues
slowly and erratically, but inexorably. The growth of very high bandwidth
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
12
communications links across the country is driven by the need for the transmission of
large quantities of data by increasingly large numbers of customers. These include
customers with science, research, commercial and consumer based needs. Especially
among the latter are those needs arising from the entertainment industry. We project that
the collective demands of these users will ensure the development of a basic affordable
capability to accomplish universally available networking.
At the present time, however, these concepts are still quite immature. Although,
conceptual standards have been adopted, the situation is still quite chaotic. Networking
is passing through a Tower of Babel building period where communications between
different entities invariably have major complications. Each manufacturer generally
solves his own product line problems reasonably quickly and reasonably well. However,
communication between units of disparate manufacture is frequently an unsolved
problem. It may very well take to the end of the century to bring about a level of
universality of networked computer communications approaching that of our telephone
system today.
As communication problems are solved, one can expect a coalescence of some of the
niches which have developed heretofore. Certainly, there will continue to be very
powerful computing engines, the supercomputers, for solving the most demanding
computational problems. Also, there will continue to exist mainframes appropriate to
massive data processing and the management of large files whether centralized or
distributed. The largest of each of these systems will certainly rely upon parallel
processing approaches to attain the required levels of performance. They may or may not
utilize vector processing approaches as well. What the architecture might look like is a
question that only the successful conclusion of currently active research and product
development programs will answer. Whether the level of parallelism is only modest (a
few to a few hundred processors) or where the processor count is in the thousands is an
open question. Much progress must be made, especially in the software area, before
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
13
these questions can be settled. For a long time, however, it is likely that a wide range in
the level of parallelism will be explored.
Because of their cost niches, minisupercomputers and also minicomputers are likely to
continue to be identifiable market niches. However, the demarcation between the
minicomputer and the applications workstation and possibly between the minicomputer
and the larger mainframes may very well disappear.
Good networking and good inter-unit communications will make the breadth of options
quite wide. Which options are utilized will be determined by: cost, the nature of the
problem, the size and type of databases accessed, organizational division lines, and, as
always, history.
Especially for the scientific and engineering communities, very powerful desktop
graphics will certainly play a major role. For almost all users, graphics (including color)
will play an important role. Also, the user friendly tools that have become the norm for
the personal computer will continue to be extended and span almost all human/computer
interfaces.
Applications software will continue to evolve explosively. The user will become further
and further relieved of understanding the working innards of the computer. Which
computer in a network is being used will become of less interest to the user; possibly
knowable only on explicit demand. Higher level languages will continue to evolve to
accomplish these objectives.
Ideally, what a user wants in any discipline is a nonobtrusive desktop device through
which requests or inquiries may be made and responses received in as short a time as
possible. Ideally, the workstation or the device on the users desk would be inexpensive,
generate little noise and heat, take up little space on the desk, and would yield output as
required, and make hard copy inobtrusively and quickly. Whether the actual processing
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
14
components are in the box, down the hall, or in a central location some distance away
(possibly across the country) , should be of no consequence to the user.
What will the computer spectrum of 1993 or 2000 look like? It is safe to predict that the
trends of the last 30 years will continue. Niches will continue to evolve and revolutions
in thought will continue to be ameliorated by the pragmatic realities of the marketplace.
As Toeffler warned, dramatic change has itself become a stable process. Nowhere is this
more evident - and more satisfying - than in the computer spectrum.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
15
NOTES AND REFERENCES
1. Alan Toeffier, Future Shock, (Random House, New York, NY, 1970).
2. The internal struggles within IBM that preceded the decision to commit to the System
360 as a single product line were enormous. The story is well documented, see
e.g., Pugh, E.W., "Memories That Shaped an Industry" (MIT Press, Cambridge, MA,
1984); also the articles in Fortune magazine, September 1966, p. 118 and October
1966, p. 140.
3. Amdahl, G.M., Blaauw, G.A., Brooks, F.P. Jr., IBM Journal of Research and
Development, 8, 87-101 (1964); Amdahl, G.M., Blaauw, G.A. Brooks, F.P. Jr.,
Padegs, A., Stevens, W.Y., IBM Systems Journal, 3, 119-261 (1964).
4. Evans, B.O., "System/360: A Retrospective View," Annals of the History of
Computing (AFIPS Press, Chicago, IL, 1986).
5. A definition of supercomputer(s) as given in Supercomputing,,An Informal Glossary
of Terms, prepared by the Scientific Supercomputer Subcommittee of the Committee
on Communications and Information Policy (Institute of Electrical & Electronics
Engineers, Inc., New York, NY, 1987): "At any given time, that class of general-
purpose computers that are both faster than their commercial competitors and have
sufficient central memory to store the problem sets for which they are designed.
Computer memory, throughput, computational rates, and other related computer
capabilities contribute to performance. Consequently, a quantitative measure of
computer power in large-scale scientific processing does not exist and a precise
definition of supercomputers is difficult to formulate."
6. Programmed Data Processor -1 Handbook (Digital Equipment Corp., Maynard, MA,
1960).
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
16
7. IBM 650 DP Systems Bulletin, "General Information, Console Operations, and
Special Devices," G24-5000-0 (IBM Corp., 1958).
8. Riganati, J.P., Schneck, P.B., Computer, 17, 97-113 (1984).
9. Control Data Corporation, Control Data STAR-100 Computer, St. Paul, Minnesota,
1970.
10. Cray, S.R. Jr., "Computer Vector Register Processing," United States Patent, No.
4,128,880 (1978).
11. Barnes, G.H., Brown, R.M., Kato, M., Kuck, D.J., Slotnick, D.L., Stokes, R.A.,
IEEE Transactions on Computers, C-17, 746, (1968); Boulcnight, W.J., Denenberg,
S.A., McIntyre, D.E., Randal, J.M., Sameh, A.H., Slotnick, D.L., Proceedings of
the IEEE, 60, 369, (1972); ILLIAC IV Systems Characteristics and Programming
Manual, Burroughs Corporation (1972).
12. Hockney, R.W., Jesshope, C.R., Parallel Computers (Adam Hilger Ltd, Bristol,
1981).
13. For a recounting of the early personal computer story, see e.g., Rogers, E.M.
and Larsen, J.K., "Silicon Valley Fever" (Basic Books, Inc., New York, NY, 1984); .
Freiberger, P. and Swaine, M., "Fire in the Valley" (Osborne/McGraw-Hill, Berkeley,
California, 1984).
14. Although "unit cost" and "annual dollar revenue" may be described in quantitative
terms, these numbers are difficult to obtain with precision.
There are questions such as whether or not the cost of peripherals and software are
included. Indeed, is a disk - necessary for a system to function - included as
part of the system or as a separate peripheral? Also, the demarcations between
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
?
some of the categories, e.g., minicomputers and mainframes, are likely to be
made differently by different manufactures. Thus, the more quantitative axes also
must be interpreted as having quite large uncertainties.
15. The primary data for generating this figure is from Dataquest. Additional and
corroborating data was obtained from Hambrecht and Quist, and the Gartner Group.
16. A report on this subject is in preparation by the IEEE Scientific Supercomputer
Subcommittee.
Acknowledgements: We would like to acknowledge the assistance of Neil Coletti for the
graphical presentation, Diana Evans for the manuscript
preparation, Linda Bringan for library assistance in the research
and Heidi James, the IEEE staff member responsible for
innumerable activities that made this paper possible.
17
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
18
?
FIGURE CAPTIONS
? FIGURE 1. The changing profile of computing space, over the last three decades: (a)
1960; (b) 1970; (c) 1977; (d) 1982; (e) 1987. The Annual $ Revenue peak represents
total sales worldwide in dollars current for the year listed. See Reference (15) for sources
of data. Note that log scales are used for the Unit Cost and Annual $ Revenue axes.
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4
3108 ?
SIB ?
COMMERCIAL
COMPUTERS
(S.371 BI
SCIENTIFIC
COMPUTERS
0.300 B)
so .111 ? ?..AiVIVANi.\-N.,?
? Ili
.111:
10003
1003
100
10
(a) 1960
S1OOB
S1OB ?
ii1, /101,,
SIB ? PERSONAL 11111
COMPUTERS
0.086 B) LI,' 40 1044 SUPERCOMPUTERS
(9.021 8)
SO.IB ? Afittwilk eit/zoN\,,
?rt--.4*7walik:
MAINFRAMES
MINT (5992 9)
COMPUTERS
(13.13 B)
10
10000
(c) 1977
ANNUAL S
REVENUE.
$1008 ?
S1OB ?
SIB ?
60.113
COMMERCIAL
APPLICATIONS
PERSONAL
COI.MUTERS
(3331 B)
MADIFIAMES
COMPUTERS
(531.2 B) (1233 B)
MINISUPER?
COMPUTERS
(1.133 R)
SUPERCOMPUTEAS
(5910 II)
SCIENTIFIC
APPUCATIONS
IOC)
UNIT COST (S000)
10000
(e) 1987
Declassified in Part - Sanitized Copy Approved for Release 2012/08/16: CIA-RDP90G01353R001100170002-4