rick@cs.arizona.edu (Rick Schlichting) (02/26/91)
[Dr. David Kahaner is a numerical analyst visiting Japan for two-years
under the auspices of the Office of Naval Research-Asia (ONR/Asia).
The following is the professional opinion of David Kahaner and in no
way has the blessing of the US Government or any agency of it. All
information is dated and of limited life time. This disclaimer should
be noted on ANY attribution.]
[Copies of previous reports written by Kahaner can be obtained from
host cs.arizona.edu using anonymous FTP.]
To: Distribution
From: David K. Kahaner ONR Asia [kahaner@xroads.cc.u-tokyo.ac.jp]
Re: Japanese computing report from S. Jarp CERN [sverre@cernvm.cern.ch]
23 Feb 1991
ABSTRACT.
A discussion of high-end Japanese computing with needs of European high
energy physics community in mind is presented.
INTRODUCTION AND SUMMARY.
Mr. Sverre Jarp spent the summer of 1990 in Japan in order to study
Japanese computers' relevance to the European community and to his
laboratory (CERN). Jarp is very well qualified in this regard, as he is
the Software Group Leader in the Data Handling Division of CERN, the
European Laboratory for Particle Physics, in Geneva. Jarp was given an
office at IBM's Tokyo Research Lab and used that as a base for trips to
other laboratories and to vendors. I had several lengthy and insightful
discussions with him during this time. Subsequently he has written a
important report on his visit which he has given me permission to
distribute (below). His main conclusions are as follows.
* Japanese computer industry has mostly satisfied domestic demand and
is now embarked on an effort to compete successfully in the export
market.
* Japanese component technology has been state-of-the-art for some time.
Silicon improvements are still being pursued although several newer
technologies are quite possible.
* A year or two will be needed before the latest Japanese hardware,
configurations, Unix systems, and applications will be available.
* By 1993 Japanese computer vendors should have about the same proportion
of the world-wide computer market (42%) as American vendors.
* Europe should encourage competition between American and Japanese
vendors, leading to lower prices, more software, and supplementary
economic benefits for the general European community.
* Europe should build its software industry as rapidly as possible to
benefit from portability and open systems.
* Europe should master system integration rather than producing individual
hardware elements.
* Jarp is skeptical about direct European competition with American and
Japanese electronics firms in hardware systems.
Readers of reports that I have written in Tokyo will find that Jarp's
report contains many additional and helpful perspectives. I agree with
his conclusions about Japan, perhaps emphasizing more the importance of
the Japanese consumer electronics industry as the fuel for a growing
computing industry. I am not in a position to comment about Europe.
Jarp's report also contains appendices, giving some specifications of
the latest supercomputers from Hitachi, Fujitsu, and NEC. Other
appendices list supercomputers in Japan. These are especially difficult
to keep up to date and to insure their accuracy. All such lists mostly
use the same sources, including the US Department of Commerce,
newspapers, private researchers, etc. The main differences are date of
publication and arrangement. I have checked Jarp's list and feel that
it is accurate. It essentially agrees with the one I distributed last
year although it is organized rather differently. The report's sections
are as follows.
1 INTRODUCTION
1.1 Japan's pursuit of economic power and technological leadership.
1.2 A vigorous computing industry
2 JAPAN'S DEMOGRAPHY AND GEOGRAPHY.
2.1 Relations with the West. Demography and geography
2.2 Does Japan Inc. Exist?
2.3 Exchange with the Western World
3 THE JAPANESE ELECTRONICS INDUSTRY
3.1 Japanese electronics companies
3.2 Overview of Fujitsu, Ltd.
3.3 Overview of Hitachi, Ltd.
3.4 Overview of Nippon Electric Company (NEC)
4 SUPERCOMPUTER EVOLUTION
4.1 Introduction
4.2 Initial development of Japanese supercomputers
4.3 High-end development philosophy
4.4 Installed base history
4.5 Comparison between Japan and Europe
4.6 CRAY in Japan
4.7 IBM in Japan
5 CURRENT JAPANESE SUPERCOMPUTERS AND MAINFRAMES
5.1 Introduction
5.2 NEC's SX series
5.3 Fujitsu's VP series
5.4 Hitachi's S-series Overview
5.5 Japanese mainframes
6 JAPANESE SYSTEM SOFTWARE AND COMPILERS
6.1 Proprietary systems
6.2 Manufacturers' involvement with UNIX
6.3 UNIX usage in Japan
6.4 UXP/M and Super-UX
6.5 Assemblers and compilers
6.6 Application software
7 SUMMARY AND CONCLUSIONS
7.1 The Japanese computer industry
7.2 Future evolution
7.3 Implications for Europe
Appendix A NEC SX-2A and SX-3 model characteristics
Appendix B Fujitsu VP-400E and VP-2600 model characteristics
Appendix C Hitachi S-810 and S-820 model characteristics.
Appendix D CRAY supercomputers in Japan
Appendix E ETA supercomputers in Japan
Appendix F Fujitsu supercomputers in Japan
Appendix G Hitachi supercomputers in Japan
Appendix H NEC supercomputers in Japan
------------------------------------------------------
A REVIEW OF JAPAN AND JAPANESE HIGH-END COMPUTERS.
Sverre Jarp
Computing and Networks Division
European Laboratory for Particle Physics (CERN)
1211 Geneva 23 Switzerland
Tel: +41-22-767-49-44, or: +41-22-767-33-49
Fax: +41-22-767-71-55
Email: sverre@cernvm.cern.ch, or: sverre@cernvm.bitnet
25 January 1991
ABSTRACT
--------
The author spent two months in Tokyo during the summer of 1990 studying
the Japanese high-end computing scene with the aim of understanding how
CERN (and the West in general) may be affected by the increasing
strength of the Japanese computer companies.
Several large Japanese manufacturers already have very interesting
products on the hardware side, both general mainframes and
supercomputers. Today, they claim to have the fastest systems available,
and they are working on future technologies like Ga-As or Josephson
junctions to further increase their strength.
Up until now, however, the software and communications facilities have
not been the strong point of the Japanese offerings. The issue is
therefore to try to establish whether the standardisation of system
software (via UNIX) and communications infrastructures, as well as the
globalisation of the market place, will allow them to move in rapidly as
a major supplier of high-end computing equipment.
The main part of the report is made up of descriptions of the latest
supercomputer systems offered by NEC, Fujitsu and Hitachi, as well as a
brief overview of their operating systems and current installed base. A
reference is made to the systems installed by IBM and CRAY in Japan. A
comparison is made to the European supercomputer installed base. The
fact that supercomputers were chosen is merely tactical, but these
systems do convey very clearly the technological strength of a
manufacturer since supercomputers are always built with state-of-the-art
technology. The first two chapters of the report are aimed at providing
background material for understanding Japan as a nation. The conclusion
tries to predict what will happen in Europe in the large-scale computing
area over the next few years, and offers the opinion of the author on
how best to profit from the situation.
DISCLAIMER
----------
Clearly this report (as every other report) reflects the author's
background and experience. I have tried to absorb enormous amounts of
information related to the topic, but the Japanese computing scene with
its ramifications is broad and complex. Facts have been stated to the
best of my awareness (and corrections are welcome). I have tried to be
objective and to look at the Japanese computer manufacturers in
comparison to others with neutral eyes. I have no illusions, however,
that I have not always succeeded. The report is biased by the fact that
supercomputer hardware was chosen to analyse the Japanese computing
scene to the detriment of mainframes and/or an in-depth analysis of
application software packages and their individual performance
characteristics on selected platforms.
ACKNOWLEDGEMENT
---------------
This study was made possible due to support from IBM Switzerland, IBM
Europe, and CERN. The author was also assisted by CRAY Switzerland as
well as IBM Japan, CRAY Japan, Fujitsu Ltd., Hitachi Ltd, and Nippon
Electric Company (NEC). A large number of good friends around the world
should also be thanked.
CHAPTER 1. INTRODUCTION
=======================
1.1 Japan's pursuit of economic power and technological leadership
---------------------------------------------------------------------
Japan with only 123 million people, half of the population and one
twenty-fifth the geographical size of the United States, has now risen
to become an economic giant. Although the US Gross National Product is
still almost twice as big as that of Japan (5 trillion dollars against
2.7 trillions in 1989), the growth of Japan's GNP is higher and the GNP
per capita overtook that in America a few years ago; additionally the
Japanese economy is extremely solid, with a trade surplus of 65 billion
dollars annually, whereas the US has a deficit of a similar size ( $ 50
billion in 1989) Japan has therefore become a country with formidable
financial resources to back her broad electronics industry and her other
critical industrial sectors.
This report will not go deeply into the economic or the resulting
political issues and the reasons behind them. What will be analysed in
some depth, however, is the Japanese producers of large-scale computers,
both mainframes and supercomputers, and try to indicate what their
strengths or weaknesses are.
The future of the vital domain of computer technology is certainly going
to be decided between the triangle formed by the US (or the North
American continent), the European countries, and the Pacific Rim (mainly
Japan, but also Korea to some extent) over the next few decades. The
reason for including Europe is the fact that after 1992 Europe will be
the biggest single market in the world, and the belief that this will
not only stimulate Japan and the US, but also the (embryonic) European
computer industry itself. It is hoped that this report will help people
in Europe understand Japan's success in building up a computer industry
that is today both broad and strong.
About 30 years ago, Japan was entirely dependent on American computing
equipment (mainly made by IBM). To reduce their dependency on foreign
systems, they employed a scheme of strong governmental support for
reproducing other companies' computer designs and began to build up
their own computer industry. For about two decades they were largely
content to produce mainframes that imitated the IBM/370 systems, selling
them in the local market with a Japanese MVS (Multiple Virtual Systems)
look-alike operating system, or as plug-compatible machines abroad. The
Japanese manufacturers would typically wait for IBM to make a new
announcement, and then announce 'compatible' systems within a given time
delay.
In recent years we have seen several changes in this scenario:
* First of all the Japanese companies now dominate the domestic market
by providing an impressive 80 % of all computer equipment. Japanese
technology is now state-of-the-art, and this has led to a self-assurance
amongst the Japanese where they no longer wait for American vendors to
announce in order to come out with similar products.
* The large Japanese manufacturers have expanded beyond mainframes into
supercomputers and are today claiming to have the fastest supercomputers
on the market.
* UNIX is about to remove the burden of having to stay compatible with
IBM's proprietary operating system and should give the Japanese a
greatly improved marketing platform for selling their systems in Europe
and in the US.
This report will look in detail at the three largest computer
manufacturers (Fujitsu, Hitachi, and NEC) and their high-end computer
products. A quick review of the situation of CRAY and IBM in Japan will
also be undertaken. It was beyond the scope of this report to cover the
other partners in the broad computer (and electronics) industry in Japan
as well as new development areas like massively parallel computing.
1.2 A vigorous computing industry
---------------------------------
The best demonstration of the vigour of the computer manufacturers is
the list of events that occurred at around the same time as the author's
visit to Japan (July - August 1990):
May: Fujitsu's new supercomputers start shipping (VP-2000 series).
June 12th: Hitachi announces new mainframes series (4-way M-880).
July 4th: NEC announces new mainframe series (6-way ACOS-3800).
July/Aug: Fujitsu announces ICL purchase.
August: NEC SX-3 available for benchmarking.
Aug. 30th: Fujitsu announces new VP models, new version of UNIX.
Sept. 4th: Fujitsu announces new mainframe series (8-way M-1800)
In particular the last announcement which happened "co-incidentally" the
day before IBM made its biggest announcement in 25 years, can probably
be taken as a clear sign from the Japanese that they want to be
considered to be in a position of technological leadership.
Chapter 2 Japan's Demography and Geography.
===========================================
2.1 Relations with the West. Demography and geography
-----------------------------------------------------
As already mentioned Japan has 123 million inhabitants living on about
372 thousand square kilometres. This corresponds to a population density
(331 inhabitants/square km) which is one of the highest in the world
after Bangladesh, South Korea and Holland. The geographical size,
however, corresponds to that of Finland in Europe or the state of
Montana in the US. (The population in Montana is less than one million
inhabitants). Japan consists of 4 big islands (Hokkaido, Honshu,
Shikoku, and Kyushu), and innumerable smaller ones. The main population
lives on Honshu (78%), followed by Kyushu (11%), Hokkaido (5%), and
Shikoku (4 %). The Japanese islands have always been rather hostile due
to climatic and meteorological hardships. For this reason the population
inhabits the plains that used to provide (and still provide) the best
rice-growing capabilities. A large proportion of the Honshu population
is consequently to be found in the Kanto plain surrounding Tokyo or the
Kansai plain surrounding Osaka. The harsh climatic environment may be
one of the factors that have led Japanese society to adopt rules for
collective behaviour that are much more rigorously enforced than in
Europe or in the US, and which seem to be at least part of the
explanation for the Japanese economic success.
2.2 Does Japan Inc. Exist ?
----------------------------
Many people wonder whether "Japan Inc." exists or not. There seems at
least to be a great deal of loyalty to the country. As mentioned earlier
this may be linked to the way the Japanese had to structure their
society simply to maximise the probability of survival on some rather
inhospitable islands. Today one finds a lot of loyalty inside companies;
the companies themselves are often faithfully integrated inside groups
(Keiretsus or others). Furthermore the government through MITI (Ministry
of Trade and Industry) is able to establish projects where there is
patriotic participation across the industry sector, allowing for
transfer of technology. The net result is a kind of a magnet which has a
strong force because its individual domains are aligned.
Some people may find this discussion misplaced, but the author believes
that Japanese economical and technological strength must also be viewed
against Japanese sociological order and behaviour.
2.3 Exchange with the Western World
-------------------------------------
One of the West's big problems vis-a-vis Japan is the fact that so few
Westerners go there, either to live or for a visit. Whereas in 1988
there were 195,000 Japanese living in North America (including Hawaii)
and 80,000 in Europe, there are only 31,000 Americans and less than
10,000 Europeans in Japan. When we consider that the European population
is about three times the Japanese one, this becomes an
under-representation of a factor of about 25 !
Altogether the Japanese are keeping more than half a million people
abroad to make sure Japanese interests are well covered in other
countries. Many Europeans also stay abroad but, as described, they are
very unlikely to be found in Japan.
Similar exchange problems can be demonstrated with tourists. Whereas 7
million Japanese went abroad in 1988, only 1 million tourists visited
Japan. Looking at the US one finds that 2.4 million Japanese visited the
US, but only 200,000 Americans 'returned' the visit. Again the ratio is
no better for Europe.
It could be argued that the lack of first-hand experience with Japan
leads to a clear lack of knowledge about the country in our society. We
do not have adequate knowledge about the country, the people, their
history, or their ambitions. Clearly both the language and the culture
in Japan are reasons for some of the unwillingness on the part of
Western people, but it is firmly believed that given the importance of
Japan in today's world, the West has no other choice than to keep itself
well informed about the Japanese. We need factual and first-hand
information about Japan allowing us to judge their strengths as well as
their weaknesses.
Finally, the Japanese are in the privileged position that they master
English well enough to be able to absorb all written material that is
issued in the West. Our problem, however, is the fact that although a
large amount of written material is issued in Japan every year, only
very small portions of this material is ever read or translated into
English (or other Western languages).
Chapter 3 The Japanese Electronics Industry
================================================
3.1 Japanese electronics companies
-----------------------------------
This report does not cover all the Japanese companies in the
semiconductor, electronics or computing field, like SONY, Matsushita,
Toshiba, or Omron, but it is important to review how broad and strong
the Japanese electronics industry really is. To that extent a list of
the world's most profitable companies has been included. Although IBM is
second in the list (after Nippon Telephone & Telegraph), the top of the
list is packed with Japanese banks (Bank of Japan, no 3; Sumitomo Bank,
no. 7; the Fuji Bank, no. 8; Mitsui Taiyo Kobe Bank, no. 10; Dai-Ichi
Kangyo Bank, no.11; the Mitsubishi Bank, no. 12; Sanwa Bank, no. 14),
followed by an impressive list of the main Japanese electronics
companies.
Hitachi 17
Matsushita Electric 24
Toshiba 40
NEC 50
SONY 61
Fujitsu 71
Nintendo 89
Mitsubishi Electric 95
SHARP 105
Sanyo Electric 106
In contrast, Europe's electronics industry can only point to Siemens
(46) and BASF (158). Similarly, the United States (apart from IBM) can
only point to HP (128), DEC (130), and Motorola (138). Clearly the
Japanese success in computers is related to the success and the strength
of the electronics and semiconductor industry itself.
Additionally it seems important that the Japanese government through
MITI has managed to get the large companies to line up behind government
projects like the VLSI project, the 5th generation project (with the 6th
coming), the 10 Gflop supercomputer project and now the Image-processing
(Ga-As) project and others. The fact that the resulting implementations
of a product can then be fairly similar across the participating
companies, does not surprisingly seem to be a problem for the
participants.
The next sections will review NEC, Fujitsu, and Hitachi. All three
companies are very large (in terms of annual sales) and they are all
extremely active in the semi-conductor field and telecommunications, as
well as in the field of general purpose computers and more recently
supercomputers. All three companies enjoy a healthy annual growth, both
in terms of revenue and number of employees.
3.2 Overview oF Fujitsu, Ltd.
-----------------------------
Founded in 1935 Fujitsu is the 'youngest' of the three Japanese computer
giants. It specialises in three fields; Information Processing (IP)
which covers computers, peripherals and others; Telecommunications (TC)
which covers switching systems and transmission systems; and Electronic
Devices (ED) which covers semiconductors and electronic components.
By comparing computer sales in Japan, one discovers that Fujitsu is the
market leader with sales of about 1.4 trillion yen (approximately 14
billion Swiss francs) in 1989, ahead of IBM with 1.19 trillions, NEC
with 1.17 trillions, and Hitachi with 0.92 trillions. Fujitsu also had
the highest growth rate with 14.6%. Amongst its divisions IP represents
66% of total sales, TC 16% and EC 14% (with 4% other activities). The
Fujitsu Group (including all subsidiaries) had 104,500 employees in FY89
and expanded to 115,000 employees in FY90. Total income amounted to 2.35
trillion yen in FY89 and 2.55 trillions in FY90.
3.3 Overview of Hitachi, Ltd.
-----------------------------
Started in 1910 as an electrical repair shop for a copper-mining company
in Japan, the Hitachi Group is now the 17th largest company in the world
with five diversified operating divisions (Power Systems and Equipment;
Consumer Products; Information and Communication Systems and Electronic
Devices(ICS/ED); Industrial Machinery and Plants; Wire and Cable,
Metals, Chemicals, and other Products). Total sales in FY89 were 6.38
trillion yen and the total staff was 274,000. In FY90 sales were 7.08
trillions and the staff had grown to 290,800.
ICS/ED produces a very broad range of equipment including computers,
computer terminals and peripherals, workstations, magnetic disks,
Japanese word processors, Telephone exchanges, Facsimile equipment,
Broadcasting equipment, Integrated circuits, semiconductors, picture
tubes, CRT displays, Liquid crystal displays, magnetrons, test and
measurement equipment, analytical instruments, medical electronics
equipment. ICS/ED represents 33% of the total sales.
3.4 Overview of Nippon Electric Company
---------------------------------------
Founded in 1899 as an importer and manufacturer of communications
equipment such as telephone sets and switching equipment, NEC belongs to
the Sumitomo group (Keiretsu), and used to be called the Sumitomo
Electric Company. Today NEC is ranked 39 in the world with five branches
to its corporate NEC-tree; Computers and Industrial Electrical Systems
(43 % of total sales); Communications (26 %); Electrical Devices (19 %);
Home Electronics (7 %); New Opportunities (5 %); all of it based on a
solid technology foundation.
Total sales in FY89 were 3.1 trillion yen and the total staff was
104,000. In FY90 sales were 3.3 trillions and the staff had grown to
115,000. It is the world's leading producer of semiconductors (ahead of
Toshiba and Hitachi), one of the largest producers of telecommunications
equipment (half the size of AT&T), and fourth computer manufacturer
(behind IBM, DEC and Fujitsu).
Chapter 4 Supercomputer evolution
==================================
4.1 Introduction
----------------
Less than a decade ago there were no Japanese supercomputers. The first
models were announced in 1983. Naturally there had been prototypes
earlier (like the Fujitsu F230-75 APU produced in two copies in 1978)
but Fujitsu's VP-100 and Hitachi's S-810 were the first officially
announced versions. NEC announced its SX-1 slightly later.
The last seven years seem to have been hectic. Two generations of
machines have been produced by each manufacturer and model improvements
have also been offered during the life-span of these machines. During
seven years about 150 systems have been installed in Japan (with
relatively few installations outside the country), and a whole
infrastructure of supercomputing has been established. All major
universities have supercomputers, most of the large industrial companies
and research centres as well; and there are well established
supercomputer research institutes and industry observers.
4.2 Initial development of Japanese supercomputers
--------------------------------------------------
Based on their own success with mainframes and the success of the CRAY-1
and the CDC Cyber 205, the Japanese decided in the late seventies to
start producing vector-based supercomputers. The first versions were
rather primitive, but in a short period of time all three manufacturers
have gone from basic implementations to versions that today are
considered to be amongst the best in the world. This is a remarkable
achievement and underlines the fact that a company that possesses the
underlying technology can relatively easily progress to the point of
mastering logic design and the ensuing implementation complexity.
The Japanese supercomputers were initially orientated towards parallel
pipelines featuring multiple floating-point units always governed by one
control processor. With the advent of the latest generation of systems
from Fujitsu and NEC these supercomputers have added the dimension of
multiprocessing; Hitachi's next system is bound to do the same. The
change logically follows from the fact that only computing problems with
long vectors and the right mix of floating-point instructions could
expect to move towards peak performance in the early versions of these
supercomputers. Multiprocessing adds a dimension of versatility to the
hardware, but the price to pay is added complexity in the software
requiring the basic operating system, the compilers, the libraries, and
the applications to be made aware of this architectural feature. In
contrast, the design of the CRAY X-MP took almost the opposite approach,
with multiple CPUs, each featuring only one add and one multiply
single-pipeline functional unit . It is expected that in the future all
systems must include both multiprocessors as well as multiple pipelined
functional units per processor to remain competitive. We should
therefore expect the Japanese to increase multiprocessing to 8, 16, or
more; and Cray and others to offer more pipelining per CPU. According to
Cray Research their next machine, the C-90, will offer 16 CPUs, with two
sets of pipelined add and multiply floating-point units per CPU.
All three Japanese manufacturers currently employ the scheme of driving
the vector processor at half of the scalar cycle time (by using a minor
cycle time and a major cycle time). CRAY-2 employed a similar scheme of
minor cycles, but both the CRAY X-MP and Y-MP employ full cycles.
Whereas this is good for the vector performance, it can lead to an
aggravating imbalance in the total system, since the scalar Flops are
orders of magnitudes lower than the vector Flops. Such systems can run
the risk of being suitable for only a narrow set of applications,
demanding general-purpose systems for the applications that do not
vectorise well. In other words they promote the classical front-end
back-end combination.
4.3 High-end development philosophy
-----------------------------------
Whereas the original supercomputers in Japan were developed from the
existing mainframes by adding vector processors, a decade later the
Japanese have now moved into a position where they concentrate the
development on supercomputers and obtain the mainframe computers as a
'by-product' by employing the scalar processor of the supercomputer as
the general-purpose processor in the mainframes. Clearly the memory
system has to be somewhat redesigned, but the technology remains the
same. This philosophy allowed NEC recently to announce the ACOS/3800
6-way mainframe processor based on the SX-3 and Fujitsu to announce the
M-1800 8-way system based on the VP-2000 processor design.
Financially this approach must be very attractive. It allows the R & D
needed to develop competitive supercomputers to be amortised not just
over a few dozen of these, but rather over several hundred
supercomputers and mainframes combined.
Vis-a-vis other manufacturers, CRAY in particular, this seems to be a
strong competitive advantage for the Japanese and may provide one
explanation of the apparent cost-effectiveness of their hardware
systems. The philosophy of IBM and DEC of expanding their mainframes to
super-computers by the addition of a vector facility should offer a
similarly attractive cost advantage.
4.4 Installed base history
--------------------------
Several lists exist of supercomputer installations in Japan. This report
is based on an updated version of a list circulated by D. Kahaner
(ONRFE, Tokyo). Appendices D - H list detailed installations for each
manufacturer. By analysing the installation dates it can be shown that
up until 1983 there were only two supercomputers in Japan (both
Cray-1s). In 1983, as already stated, the first domestic systems were
delivered, but until the end of 1985 there were less than 30 systems
installed. 1986 - 1988 then provided three 'golden' years with about 30
super-computers installed each year. Although the lists available may be
incomplete concerning the last two years, there is strong evidence to
conclude that the Japanese market has now come to a point where market
growth will definitely be lower.
1980 2
1983 3
1984 8
1985 13
1986 33
1987 29
1988 31
1989 16
1990 16
Total 151
Table 1 Systems installed by year in Japan.
-------------------------------------------
4.5 Comparison between Japan and Europe
---------------------------------------
Although the aim of this report is to review high-end computing in
Japan, it may be instructive to compare Europe and Japan in terms of
supercomputer installations. In 1990 the Japanese distribution was
thought to be the following (both excluding and including vendors' own
installations).
Company Manufacturer exclusive Inclusive
Fujitsu 63 73
Hitachi 18 29
CRAY 26 26
NEC 18 21
CDC/ETA 2 2
Total 127 151
Table 2 Japanese supercomputers by vendor.
------------------------------------------
A certain subset of IBM's 3090 should be added to these numbers. The
Department of Defense (DoD) rules for export of supercomputers would
suggest the addition of all 3090 systems (model 180 or higher) with at
least one vector facility and all 3090-600S or 600J (VF or not). It is
estimated that about 65 such systems exist in Japan, but it was not
possible to establish a detailed list of IBM installations for the
verification of this number. The number of Vector Facilities is
estimated to be about 110 .
In Europe the supercomputer distribution is quite different according to
recent estimates:
CRAY 71
Siemens/Fujitsu 10
Amdahl/Fujitsu 7
CDC/ETA 6
NEC 2
Total 96
Table 3 European supercomputers by vendor.
------------------------------------------
The IBM systems (DoD-categorised 3090s) exceeds 100. One noticeable
difference is the European Academic Supercomputing Initiative (EASI)
that promoted 3090-600 systems (with 6 VFs) in several European Academic
Institutes. A total of 18 EASI sites are in existence. They are RWTH
Aachen (G), CEA Paris (F), CERN Geneva (CH), CINECA Bologna (I), CIRCE
Paris (F), CNUSC Montpellier (F), DESY Hamburg (D), ETH Zurich (CH), FCR
Barcelone (E), GSI Darmstadt(D), IN2P3 Lyon (F), KFK Karlruhe(D), KUL
Leuven (B), Rome University (I), SARA Amsterdam (NL), Vienna University
(A), TU Braunschweig (D), UMEA Skelleftea (S). Adding RAL (UK), IBM
Bergen Scientific Centre (N), and ECSEC in Rome (I) where 3090-600/VF
systems also exist one finds that scientific or university sites alone
account for more VFs than Japan has in total.
Nevertheless, the total sum of supercomputers, about 200, is the same.
This leads to the conclusion that the installed base of supercomputers
per capita is at least a factor three better in Japan than in Europe.
Given that supercomputers are basically advanced tools for industry,
research and education, this imbalance represents a real handicap for
the competitiveness of Europe.
4.6 CRAY in Japan.
------------------
As already mentioned, CRAY has operated in Japan since 1980. Today, 26
systems are installed, mainly in commercial organisations. These systems
are a mixture of CRAY-1s, X-MPs, CRAY-2s, and Y-MPs. Most of them have
now been converted to UNICOS. Relatively few systems have the largest
configuration and quite a few X-MPs or Y-MPs exist with only one or two
CPUs. There may be two reasons for this. One could be the fact that
CRAY's prices are considered high, especially given that the systems
must be purchased (most domestic systems are leased) with little or no
discount. The other explanation may be that the prime need of the
Japanese is to get access to applications that run only on the CRAY,
without a need for the maximum CPU capacity. A clear need for
application access is, for instance, demonstrated by the Japanese car
manufacturers that have acquired CRAY systems to run an application
called PAM-crash for car crash simulation.
In spite of the relatively small penetration of CRAY systems in Japan,
they are nevertheless considered as prestigious trend-setters in the
market for several reasons. The CRAY systems are architecturally very
well balanced, UNICOS is seen as a mature supercomputer operating system
and most importantly, a large number of applications (more than 600)
exist for the CRAY. With the CRAY Y-MP now installed by the Tohoku
University, which has traditionally used only NEC equipment, CRAY may
have started to penetrate seriously the academic market.
4.7 IBM in Japan.
-----------------
IBM has been present in the Japanese market for a very long time. In the
early sixties, when the Japanese government decided to react to foreign
dominance, IBM controlled about 80% of the market. Although the Japanese
have reversed this situation IBM still plays a very important part in
the Japanese computing scene. In 1989 IBM sold equipment worth 1.19
trillion yen (about $ 8.5 billion, equivalent to approximately 15% of
IBM's world-wide sales).
The commercial companies, in particular, seem to be large IBM customers.
It is not unusual to see computer centres with several 3090 systems
installed (all running MVS) and huge DASD farms with hundreds of
Gigabytes. In total it is believed that several hundred 3090 systems
exist in Japan. The commercial companies seem to appreciate IBM for
their total system integration and well-balanced systems. IBM is able to
offer a wide range of peripherals that go with their mainframes, as well
as a huge set of applications on top of MVS, both from IBM and from
third-party vendors. As already stated, IBM has not had the same
penetration in the scientific market in Japan. There has been no
equivalent to the EASI programme and the total number of Vector
Facilities is rather limited. The recently announced IBM/9000 series
will provide IBM with a more powerful system with which to compete in
the future. Each VF should offer a peak performance of 400 Mflops.
Chapter 5 Current Japanese Supercomputers and Mainframes
========================================================
5.1 Introduction
----------------
This chapter describes the current offerings from NEC, Fujitsu, and
Hitachi. A comparison is made with the previous versions. Appendices A,
B, and C contain the specifications of the supercomputer models
discussed.
5.2 NEC's SX series
-------------------
Overview
--------
The SX-3 series is the second full generation of production-level NEC
supercomputers. In 1984 NEC announced the SX-1 and SX-2 and started
delivery in 1985.
The first two SX-2 systems were domestic deliveries to Osaka University
and Sumitomo Trading Company. The SX-2 had multiple pipelines with one
set of add and multiply floating point units each. With a cycle time of
6 nanoseconds, each pipelined floating-point unit could peak at 167
Mflops. With four pipelines per unit and two FP units, the peak
performance was about 1.3 Gflops. Due to limited memory bandwidth and
other issues the sustained performance in benchmark tests was typically
less than half the peak value. For some reason the SX-1 had a slightly
higher cycle time than the SX-2 (7 ns). In addition it had only half the
number of pipelines. The maximum execution rate was 570 Mflops.
At the end of 1987, NEC improved its supercomputer family with the
announcement of the A-series which gave some improvements to the memory
and I/O bandwidth. The top model, the SX-2A, had the same theoretical
peak performance as the SX-2. A family of lower speed systems included
the SX-JA (250 Mflops), the SX-1EA (330 Mflops), and the SX-1A (665
Mflops).
NEC SX-3
--------
In 1989 NEC announced a rather revolutionary new model with several
important changes. New technology was used with logic chips that have
the highest density in industry today. The vector cycle time was halved,
the number of pipelines was doubled, but most significantly NEC added
multiprocessing capability to its new series. The new top of the range
currently features four independent arithmetic processors (each with a
scalar and a vector processing unit); and NEC has pushed its performance
by more than one order of magnitude to an impressive peak of 22 Gflops
(from 1.33 on the SX-2A). From initial benchmark results, one would
deduce that the SX-3 is now the most powerful system in the world.
The logic LSI of the SX-3 has 20,000 gates per chip and a gate switching
time delay of 70 picoseconds per gate. This is a major technological
jump from what NEC applied in the SX-2 series, namely 1,000 gates and
250 ps time delay. The packaging consists of Multi-Chip Packages (MCP)
that are made of a ceramic substrate upon which the LSIs are mounted
directly. A board is 22.5 x 22.5 cm2 and can contain a maximum of 100
LSIs. It is water cooled by the cold plate method.
The scalar unit has 128 64-bit registers. It decodes all instructions
and runs in parallel to the vector unit. It is a RISC-based design using
scalar pipelines to speed up execution. Nevertheless, the cycle time is
the full machine cycle (5.8 ns) and peak scalar Flops are roughly two
orders of magnitude lower than peak vector performance. This fact
highlights the need to push applications in the direction of full
vectorisation in order to exploit the SX-3 at its best. The scalar
processor has a 64 KB cache and a 4 KB instruction stack with 1.6 ns
access time. The cache size is no bigger than it was in the SX-2. The
processor has a sophisticated branch prediction mechanism built into the
instruction stack hardware. The instruction format is either 64-bit with
memory addresses included (for load, store and branch instructions), or
32-bit for arithmetic operations (specifying three registers). Unlike
Hitachi and Fujitsu, NEC's basic instruction set is not compatible with
that of IBM. The scalar processor supports 64-bit integers directly in
native hardware.
The vector processor is equipped (in the largest configuration) with
four pipelines integrating four floating-point units (two add and two
multiply). The compilers will optimise the vector performance
automatically by scheduling vector instructions on all the parallel
hardware, but will have to be enhanced to cope with parallel execution.
The SX-3 can cope with CRAY and IBM floating-point-format (in hardware).
IEEE formats can be expected in future systems, but not in the SX-3.
Below is a detailed overview of the various SX-3 models and their
corresponding peak performance values.
SX-3 Model 11 12 14 22 24 42 44
Arithmetic Processors (AP) 1- 1- 1- 2- 2- 4- 4-
Add/Mult pipelined units -1 -2 -4 -2 -4 -2 -4
Add/Mult FP units 4 4 4 4 4 4 4
Vector regs (KB/AP) 36 72 144 72 144 72 144
Max speed (Gflops) 1.37 2.75 5.5 5.5 11 11 22
Table 4 Model differences for the SX-3 supercomputer.
-----------------------------------------------------
Primary memory was based on 256 Kb SRAMs with 20 ns access time. NEC has
announced that this will be changed to 1 Mb memory chips in 1991. The
maximum memory configuration will then be expanded from 2 GB to 8 GB.
The total memory bandwidth is subdivided into two halves (with two
processors each) which in turn feature two vector load and one vector
store paths as well as one scalar load and one scalar store paths. Like
its predecessor, the SX-3 is probably unable to offer the memory
bandwidth needed to sustain peak performance unless most operands are
contained in the vector registers. The current maximum size of the
external memory unit (XMU) is 16 GB based on 1 Mb DRAMs with 70 ns
access time. By changing to 4 MB DRAMs in 1991 NEC will increase the
external memory to 64 GB. This is an incredible memory size. (How many
people remember 64 KB as a respectable memory size!) The system allows
eight bytes to be transferred from the XMU to memory per minor clock
cycle, giving a transfer speed of 2.75 GB/s.
There can be a maximum of four I/O processors (IOPs), each with a 250
MB/s throughput. The channels can be 3, 6, or 20 MB/s (with a maximum of
64 channels/IOP). High-speed channels operate as eight pairs of 100 MB/s
channels directly through Direct Memory Access (DMA). NEC has an
agreement with Ultranet and will provide a HPPI interface in 1991.
NEC has started shipping uni-processor versions of the SX-3 to Europe.
The University of Cologne has received a SX-3/11 and The Dutch Aerospace
Lab, NLR, will receive a SX-3/12 in May 1991. The Swiss government will
probably install a dyadic version in the second half of 1991. Four
processor versions of the SX-3 are not expected before 1992.
5.3 Fujitsu's VP series
-----------------------
Overview
--------
The VP-2000 series is the second generation of full production-level
Fujitsu supercomputers. In 1977 they produced the first supercomputer
prototype called the F230-75 APU that was a pipelined vector processor
added to a scalar processor.
In 1983 they came out with the VP-200 and VP-100 systems, which later
spun off the low-end VP-50 and VP-30 systems. In 1986 came the VP-400
(with twice as many pipelines as the VP-200) and as of mid-1987 the
whole family became the E-series with the addition of an extra
(multiply-add) pipelined floating point unit that boosted the
performance potential by 50%. Thanks to the flexible range of systems in
this generation (VP-30E to VP-400E), and other reasons such as good
marketing and a broad range of applications, Fujitsu became the largest
domestic supplier with 63 systems.
VP-2000 series
--------------
Announced in 1989, and available since March/April 1990, is the VP-2000
family with a peak performance of 5 Gflops.
Fujitsu's design philosophy (like the other Japanese manufacturers) has
been centred around the original APU design where the Vector Processor
was a distinctly separate unit from the scalar unit. Emphasis was put on
multiple pipelines with multiple floating-point units. The VP-2000
series is the first Fujitsu supercomputer with multiple scalar or vector
processors. The VP-2000 system was initially announced with four vector
performance levels (model 2100, 2200, 2400, and 2600) where each level
could have either one of two scalar processors (corresponding to a model
/10 or a mo del /20). The VP-2400/40, announced end-August 1990, doubles
the number of processors compared to the VP-2400/20, and will have a
peak vector performance similar to the VP-2600.
The following table explains the relationship of the Fujitsu models.
VP-2000 Model 2100 2200 2400 2400 2600
m.40
vector cycle time 4 4 4 3.2 3.2
vector processors (VPs) 1 1 1 2 1
scalar processors 2 2 2 4 2
vector Fl.point units 5 7 7 7 7
Mult/Add. Fl.point units 2 4 4 4 4
Pipelines/VP 1 1 2 4 4
vector regs per scalar unit (KB) 32 32 64 64 64
max speed (Gflops) 0.5 1 2 5 5
Table 5 Model differences for the VP-2000 series.
-------------------------------------------------
Like the other Japanese manufacturers, the model range is basically
constructed by removing hardware elements from the top model. Firstly
the pipelines are reduced from four to two and then to one, and finally
one of the two sets of add and multiply units is removed. The memory
pipes are reduced in a similar fashion.
The logic LSI has 15,000 gates per chip and a propagation delay of 80
ps/gate. This is a very impressive level of integration although the
corresponding NEC figures are slightly better. Both Fujitsu and NEC seem
to be at the very leading edge of VLSI today. The very high integration
in the VP-2000 series enables the entire scalar processor to sit on just
one multi-layer glass ceramic board of 61 layers, which allows
elimination of off-board signal delays for the processor. The board is
24.5 x 24.5 cm 2 and can contain a maximum of 144 LSIs.
The scalar unit has a cycle time of 6.4 ns and is connected to a 128 KB
buffer storage with an access time of 1.6 ns. This very fast
Logic-and-RAM LSI is built up of 64 Kbit chips with 3500 gates. The same
chips are used for the vector registers.
In Fujitsu's design, the vector processor sits between two scalar
processors which act as instruction processors. The vector processor can
be fed from either. Having twice as many scalar processors as vector
processors can be seen as an effort to improve the balance between
scalar and vector performance. The memory system can be configured with
2 GB of real memory using the latest LSI technology with 35 ns 1 Mb SRAM
chips. The Secondary Subsystem Unit (SSU) can have up to 8 GB of memory
using 1 Mb DRAM (100 ns) chips and Fujitsu has declared that it will
move to 4 Mb DRAMs in 1991 this allowing second-level memory systems of
32 GB.
Previous machines have been heavily criticised for the lack of memory
throughput. The VP-400 series had only one fetch/store path to memory
that ran at 4.5 GB/s. This has been improved in the VP-2000 series, but
is probably not sufficient in all cases (in particular where all
operands and the results must be fetched or replaced).
As already stated Fujitsu has been shipping the new series since April
1990. The first two VP-2600 systems were delivered to the Japanese
Atomic Energy Commission (JAERI). Via Siemens at least one system has
also been imported to Germany (a VP2400/10 at the University of
Karlsruhe). Amdahl marketed the previous version of the VP systems
(after having added MVS support). It has announced that it will not
market the VP-2000 series.
5.4 Hitachi's S-series Overview
-------------------------------
Hitachi differs from the two other manufacturers in a couple of aspects.
Firstly it does not export its supercomputers and secondly the current
offering is somewhat out-of-date compared to the latest systems from NEC
and Fujitsu. In this report the S-820 is therefore treated less
thoroughly than the other systems. Nevertheless, the S-820 should be
judged on the technology it represented at first shipment date, and
Hitachi should be judged on the technology it possesses in general. It
is believed that a new supercomputer from Hitachi will be announced
during 1991.
The S-820 series
----------------
Appendix C summarises the main characteristics of the two generations of
supercomputers manufactured by Hitachi. The S-820 system offers four
performance levels (m.20, m.40, m.60, and m.80) corresponding to the
number of pipelines per floating point unit. The lowest model has an 8
ns vector cycle time. The logic LSI has 5,000 gates per chip and a
propagation delay of 250 ps/gate. The scalar unit has a cycle time of 8
ns (major cycle time) and is connected to a 256 KB buffer storage with
an access time of 4.5 ns. This bipolar RAM is built up of 16 Kbit chips
whereas the faster LSI for the vector registers has 2,500 gates, 6.9 Kb
capacity and an access time of 2.5 ns.
S-820 Model 20 40 60 80
Vector cycle time 8 4 4 4
Mult/Add. pipeline units 3 3 3 3
Vector pipeline units 5 5 5 5
Pipelines inside unit 1 1 2 4
Vector regs per scalar unit (KB) 32 32 64 128
Data bus (8B/4ns) 1 2*1 2*2 2*4
Max speed (Gflops) 0.375 0.75 1.5 3
Table 6 Model differences for the S-820 supercomputer.
------------------------------------------------------
The memory system can be configured with 512 MB of real memory using a
technology with 20 ns 1 Mb BiCMOS chips. The Extended Storage can have
up to 12 GB of memory using 1 Mb DRAM (120 ns) chips.
Hitachi has put great emphasis on a fast memory although this has meant
limiting it to maximum 512 MB. The memory bandwidth (2 words per pipe
per vector cycle) is a respectable achievement, but it is not enough to
keep all functional units busy (if memory access is needed for each add,
multiply, and generated result). The I/O processor supports 64 channels
and half of them can be 9 MB optical channels. The total I/O capacity is
288 MB/s.
5.5 Japanese mainframes
-----------------------
As already described in section 1.2, all three Japanese manufacturers
announced new mainframe systems between July and September 1990. These
mainframes are all based on the scalar processor of the supercomputer
from that company with a higher level of multiprocessing and a different
memory system. The two-level cache is, for instance, one manifestation
of this difference.
The following table lists the latest announcements:
Mainframe Cycle time Max.CPU config. Commercial MIPS Delivery
Hitachi M-880 8.0 ns 4 way 155 4Q90
Fujitsu M-1800 6.4 ns 8 way 325 3Q91
NEC ACOS/3800 5.8 ns 6 way 375 3Q91
Table 7 Latest generation of Japanese mainframes.
-------------------------------------------------
Hitachi and Fujitsu offer their systems also as plug-compatible systems
to IBM abroad. Fujitsu offers systems via their 47% share in Amdahl who
licenses the technology and makes the systems compatible, and Hitachi
does it via Comparex, Olivetti, and Hitachi Data Systems.
NEC does not offer IBM compatible systems, but is expected to announce
UNIX-support for its ACOS/3800 in the export market (as well as
domestically). The MIPS rates are estimates of commercial MIPS. In a
scientific environment the performance is not known, but both the
Fujitsu processors and the NEC processors are estimated at about 30
Mflops (scalar) for the LINPACK 100 x 100 test.
All three vendors are expected to announce 8-way systems as the maximum
configuration of this machine generation.
Chapter 6 Japanese System software and compilers
================================================
6.1 Proprietary systems
-----------------------
As already explained, the Japanese supercomputers originally grew out of
the mainframe families. The corresponding operating systems did the
same, and since the Japanese domestic operating systems were all
inspired by IBM's MVS, these mainframe systems also invaded the
supercomputers. Fujitsu had MSP, Hitachi had VOS3, and NEC had SXOS.
The advantage for domestic installations that possessed both mainframes
and supercomputers was the 'de facto' compatibility between the two, but
both the European and the US market refused to get seriously interested
in these systems.
6.2 Manufacturers' involvement with UNIX
----------------------------------------
With the latest series of supercomputers and mainframes the Japanese
manufacturers have announced a serious interest in UNIX. Fujitsu has had
a version of UTS (UTS/M), which it obtained from Amdahl in 1985,
available on its mainframes since 1986 (native since 1987). With the
announcement of the VP-2000 series Fujitsu initially announced a VPO
(Vector Processing Option) to make UTS/M into a supercomputer operating
system, but it has now announced a consolidated UNIX offering for both
environments, 'UXP/M', which will be based on System V, release 4 and
shipped in the middle of 1991. NEC has also announced its version of
UNIX, Super-UX, for its supercomputers, not (yet?) for its mainframes.
NEC will also ship their UNIX-version in the first half of 1991. Both
manufacturers base their systems on AT&T System V and are members of
UNIX International. Hitachi has not announced UNIX for its high-end
systems, but is expected to do so with the announcement of its next
supercomputer generation. Hitachi is a member of OSF as well as UNIX
International.
In section 4.4 it was speculated that the Japanese market is
experiencing a limited growth as far as supercomputers are concerned.
This can be interpreted as an additional argument why it is vital for
the Japanese manufacturers to offer UNIX to satisfy the export market.
6.3 UNIX usage in Japan
-----------------------
In the domestic market UNIX is available in certain sectors. Workstation
systems are almost exclusively based on UNIX. The domestic market
leaders are HP/Apollo, SUN, and Sony, all with about 25 - 30% of the
market each. Beyond the workstation segment CRAY has largely converted
its customer base from COS to UNICOS. This is, of course, a marginal
UNIX penetration seen from a global market perspective, but one should
not underestimate CRAY's influence as a trend-setter in supercomputer
software. Some installations (universities and research centres) run
UTS/M as a parallel offering to the domestic MSP system, but with little
real emphasis until now. It is therefore believed that the conversion to
UNIX in the domestic market in Japan will be relatively slow and that
the Japanese manufacturers will initially target their UNIX systems to
the export market. This could imply a heavy burden on the first foreign
companies as they will have to get involved in debugging and enhancing
these versions of UNIX on large systems, in a similar fashion to the
very early customers of CRAY's UNICOS.
6.4 UXP/M and Super-UX
----------------------
Both NEC and Fujitsu have to repeat what CRAY did several years ago,
namely convert UNIX from a time-sharing system to a highly reliable and
sophisticated operating system for a supercomputer. The changes that are
necessary are rather fundamental; kernel modifications for the detailed
support of the architecture, the multiprocessor support, the memory
management scheme, the I/O subsystem, the scheduler, etc. A batch
system, NQS (Network Queuing System), has to be integrated and
significantly enhanced.
The file system needs modifications both for speed improvements, large
file sizes, and complexity. I/O drivers for the full set of peripherals
must be integrated. Reliability features need to be added to make sure
the system software can keep the machine up all the time. This is no
simple task, but a good UNIX implementation has become a requirement for
the Japanese manufacturers (at least in the scientific export market).
SUPER-UX will start shipping early in 1991 (release 1.1). It is based on
System V release 3 with many BSD extensions. In addition to the general
improvements already mentioned, it will come with a Supercomputer File
System that is implemented in parallel to the System V file system
(SVFS). It will offer support for Ethernet, FDDI, and HPPI networks
including NSC's DX and Ultranet.
UXP/M is in a similar situation. Its predecessor UTS/M + VPO (vector
processing option) started shipping in 4Q90. The file system has been
greatly enhanced with several options such as asynchronous I/O,
bufferless I/O, high-speed I/O via secondary memory, etc. Furthermore,
NQS, memory management, reliability improvements for hardware and
software as well as improved systems management facilities have had to
be integrated.
6.5 Assemblers and compilers
----------------------------
First some words about assemblers. Interestingly enough neither NEC nor
Fujitsu offered the assembler to their customers on their previous
supercomputer systems. Hitachi did offer it after domestic pressure from
the user community. Today the assembler is made available both for the
SX-3 and the VP-2000 series. For code optimisation and complex coding in
certain areas assembly programming can still be an important asset in
maximising the use of a supercomputer.
Most supercomputer programs, however, rely on a highly optimised FORTRAN
compiler. In the past the Japanese FORTRAN compilers have been optimised
for single-task vector processing. With the introduction of
multiprocessing hardware a new dimension of parallel execution has been
opened for the supercomputer users, but at the cost of complex additions
to the compiler itself. Language extensions for user-controlled
parallelism as well as automatic parallelisation techniques have to be
added. Both macrotasking at the subroutine level and microtasking at the
loop or statement level must be dealt with. NEC and Fujitsu will start
offering these capabilities as of 1991, but an additional period for
refining the techniques in light of the experience with real-life
applications in the field must be included.
6.6 Application software
------------------------
It is beyond the scope of this report to provide an in-depth review of
the application software available on the Japanese platforms. Given the
on-going effort to offer UNIX as the preferred operating system at least
abroad, it is believed that such a survey should be undertaken when the
UNIX offering is mature and the porting of applications has been carried
out on a massive level.
The Japanese manufacturers are extremely keen to be able to offer the
same applications on their platforms as CRAY. For this reason both NEC
and Fujitsu have established competence centres and collaborations in
the United States.
In certain cases there will be political pressure to stop applications
being ported to Japanese platforms. This is the case today with
PAM-crash which is a vital applications package for automobile crash
simulation and certification.
Chapter 7 Summary and Conclusions
=================================
7.1 The Japanese computer industry
----------------------------------
This report has tried to demonstrate that the Japanese seem to be
succeeding in what is thought to have been the two legs of their
national computing strategy:
* Create computing solutions (hardware and software) that will satisfy
the domestic demands as much as possible.
* Enhance or adapt these solutions so that they will compete
successfully in the export market.
The Japanese have largely fulfilled the first goal by acquiring a share
of about 80% (revenue-based) of the domestic market whereas in the
sixties they achieved a mere 20%. It is interesting to note that this
success is based on a very broadly-based computer industry as
demonstrated in Chapter 3.
On the export market Fujitsu and Hitachi have had an initial strategy to
operate as Plug Compatible Manufacturers (PCMs) to IBM. This has allowed
them to penetrate both the American market via Amdahl and HDS (Hitachi
Data Systems) as well as the European market via HDS, Comparex
(Siemens/BASF), Olivetti, and Amdahl. As long as this continues to be a
lucrative market these companies are expected to stay put. Nevertheless
the general trend to UNIX and Open Systems is believed to gradually move
the emphasis away from proprietary systems. The Japanese manufacturers
have clearly understood the importance of this shift and should offer
complete UNIX systems for both mainframes and supercomputers within the
near future (Fujitsu has already announced its UXP/M for both
environments). To secure the success of this effort the Japanese
companies have all established centres in Europe or the United States to
ease porting operations and capture new trends and evolutions in the
rapidly moving UNIX area.
Japanese component technology has been state-of-the-art for some time
already. Whether one considers chip density, switching speed or other
technological factors, the Japanese compete successfully with everybody
else in the world. With the latest series of integrated mainframes or
supercomputers the Japanese have also demonstrated that they are now
strong players who want to act as market leaders, and not as complacent
followers any longer. They have reached the stage where their integrated
products have the same complexity in terms of multi-processing or memory
subsystems as their competition. Their strategy is now to offer these
hardware systems with open system software to a world-wide market.
To create a balanced view of the situation it must be kept in mind that
some of the hardware and software features described in this report are
not yet fully available. A year or two may still be needed by the
Japanese manufacturers to be able to offer their latest hardware
systems, in their largest configurations, with a fully developed and
debugged UNIX system and a full range of applications.
7.2 Future evolution
--------------------
In the near future the Japanese will continue to enhance their systems
to improve their competitiveness in the market place. By 1993 they are
expected to have the same proportion as the Americans of the world-wide
computer market (about 42% each). NEC, Hitachi, and Fujitsu should all
by then be selling mainframes and supercomputers with solid UNIX
operating systems and a broad spectrum of applications. Follow-on models
of the existing systems can be expected, either as new families or as
upgrades within the existing families. New players in the high-end
computer market can also be expected. Matsushita has already announced
its intention to compete in the supercomputer market in the future.
Beyond today's systems the Japanese are evaluating several approaches to
improving their products. Silicon-based improvements are being pursued,
and Hitachi has already announced a laboratory version of its 64 Mbit
DRAM. Memories should therefore become larger and larger with the
advances in memory technology. All manufacturers will pursue the race
for lower cycle times (approaching the 'magic' 1 ns cycle). Because of
inherent limitations of silicon chips this race could bring out
innovative new technologies such as Ga-As or Josephson junctions, both
of which have already been explored inside the Japanese research labs
for quite some time. Integration into commercial products is believed to
depend more on what the competitors can achieve than anything else.
Fujitsu has, for instance, developed the HEMT (High Electron Mobility
Transistor) which is a variant of the Ga-As technology with promising
features both in terms of reduced heat dissipation and greatly improved
switching time. The company is already producing 64 Kb memory chips
based on HEMT and should offer integrated circuits in the near future.
Although traditional architectures will continue to dominate the market
for the next few years, the Japanese are also seriously interested in
other architectural approaches. Massively-parallel systems are believed
to be the next evolutionary step in the sophistication of their systems.
7.3 Implications for Europe
---------------------------
One of the purposes of this report was to understand the implications
for Europe of the current strength of the Japanese computer
manufacturers. Unlike the Japanese themselves, the Europeans have not
managed to build up an internal computer industry that has sufficient
strength to compete with the Americans. Europe has therefore been a
faithful acquirer of American main-frames and supercomputers (with the
notable exception of a few Japanese systems).
In 1993 Europe will be the biggest united market in the world and vital
to every large computer manufacturer that wants to succeed in the long
term. For the Old World it is critical to anticipate the implications of
this privileged position. After this study (the author has also been
involved with American computer manufacturers for the last twenty years)
it is believed that Europe should initiate a policy based on the
following principles:
* Adopt an immediate strategy of encouraging a strong and healthy
competition in the European market place between the Americans and the
Japanese computer manufacturers. This could lead to lower prices as well
as better and more varied software and hardware offerings. Manufacturers
should also be told that Europe expects supplementary benefits in terms
of local investment in factories and research and development
laboratories, which would bring additional employment opportunities and
tax money to our communities. This strategy should also ensure that
Europe becomes as well equipped with supercomputers as Japan (or the US
for that matter) and therefore maintains European competitiveness.
* Europe should build up a strong software industry as rapidly as
possible. This industry should profit from the Open Systems penetration
and build portable application packages that will satisfy not only
European demands but will also allow European software products to
compete successfully on a worldwide basis. Europe has strong traditions
in software and, although the Americans are also very strong software
builders, we can probably profit from the fact that our demands are more
complex and diversified than the those of the United States. How many
times have we experienced American software products that do not cope
with the intricate production environments in Europe? Since software is
currently also the weak point in Japan's computer strategy, we would
have an excellent chance of providing products for their systems both
domestically and abroad. This does, of course, presuppose that we become
'truly' European in our activities. If we focus solely on regional
demands and pursue only local market opportunities we will not achieve
this goal. On the other hand this strategy should not require the
existence of huge companies like the Japanese electronics giants. We
can encourage small and dynamic software houses to help us achieve this
goal.
* The third element in our strategy should be system integration. The
future of computing will be very complex. Computer manufacturers will
bring innumerable platforms to the market from hand-held micro-computers
to Tera-flop supercomputers. In addition vast numbers of peripherals,
multiple connectivity options, and evolving network protocols will all
be elements that will contribute to a high-level of complexity in our
data-processing environments, and the only realistic option is to assume
that the issue will not get simpler over time. Computer users, however,
will demand applications and systems that give a unified view of
distributed software and databases. In the author's opinion it is,
therefore, much more important to master the aspects of system
integration than to produce the individual hardware elements.
Nevertheless it assumes broad-minded companies that can evaluate the
advantages of individual computing elements and produce both a vision
and follow up the vision with a solution. The broadness of the vision
should not be the limits of Europe in 1993, it should be the limits of
the globe. Japan and its activities must absolutely be an integral part
of it on par with the US.
In contrast, the author is rather sceptical about direct European
competition with the American and Japanese electronics giants as far as
hardware systems are concerned. Siemens will, we hope, continue to be
present on the list of the world's largest companies, but up until now
it has in no way been able to initiate a computer hardware strategy
analogous to that of the Japanese companies discussed in this report. On
the contrary, the supercomputers sold by Siemens are obtained directly
from Fujitsu and the large mainframes offered by Comparex (a joint
Siemens/BASF company) are Hitachi systems. There will probably be niche
opportunities, and Europe should continually try to explore the
possibility of producing systems where added value is given to an
integrated computer product even if the components are largely bought
off the shelf in Japan or the US.
In recapitulation:
------------------
Up until now, computers were mainly supplied to Europe by the US. In the
future they should be acquired from both the US and Japan. Rather than
hoping for Europe to become a computer supplier of the same calibre as
these two, we should exploit fully this competitive situation as well as
the opportunities for providing value-added software solutions and
highly qualified system integration.
Appendix A NEC SX-2A and SX-3 model characteristics.
====================================================
Table 1 Main features of the last two generations of NEC supercomputers
-----------------------------------------------------------------------
System SX-2A SX-3
Scalar processors 1 4
Scalar cycle time (ns) 6 5.8
Vector processors 1 4
Vector cycle time (ns) 6 2.9
Gates in logic 1,000 20,000
Switching time (ps) 250 70
Cache technology/chip density 1 Kb bipol.RAM 40 Kb +7K gates
Cache access time (ns) 3.5 1.6
Cache size (KB/CPU) 64 64
Vector regs. (KB/CPU) 80 144
Floating-point units 2 (Mult/Add) 4 (2 Mult/Add)
Pipelines per F.P. unit 4 4
Max Gflops 1.3 22
Memory interleave 512 way 1024 way
Memory transfer rate (GB/s) 11 80
Main memory technology 256 Kb SRAM 256 Kb SRAM
Memory access time (ns) 40 20
Max. memory (GB) 1 2
Second-level memory 256 Kb SRAM 1 Mb DRAM
Access time (ns) 70
Max. memory (GB) 8 16
Transfer rate to main (GB/s) 1.3 2.75
I/O units 1 4
Total I/O throughput 192 MB/s 1 GB/s
Initial shipment date June 88 Sept.90
Appendix B Fujitsu VP-400E and VP-2600 model characteristics.
=============================================================
Table 1 Main features of last two generations of Fujitsu supercomputers
-----------------------------------------------------------------------
System VP-400E VP-2600
Scalar processors 1 2
Scalar cycle time (ns) 14 6.4
Vector processors 1 1
Vector cycle time (ns) 7 3.2
Gates in logic 400/1,300 15,000
Switching time (ps) 350 80
Cache technology/chip density 4 Kb 64 Kb
Cache access time (ns) 5.5 1.6
Cache size (KB) 64 2*128
Vector regs. (KB) 128 2*128
Floating point units 3 (Add+M-Add) 4 (2 Mult-Add)
Pipelines per F.P. unit 4 4
Max Gflops 1.7 5
Memory interleave 128/256 way 512 way
Memory transfer rate (GB/s) 4.5 20
Main memory technology 256 Kb SRAM 1 Mb SRAM
Memory access time (ns) 55 35
Max. memory 256 MB 2 GB
Second-level memory 256 Kb SRAM 1 Mb DRAM
Access time (ns) 100
Max. memory 768 MB 8 GB
Transfer rate to main (GB/s) 4.5 10
I/O units 1 1
Total I/O throughput 96 MB/s 1 GB/s
Initial shipment date Dec . 87 April 90
Appendix C Hitachi S-810 and S-820 model characteristics.
=========================================================
Table 1 Main features of two generations of Hitachi supercomputers
------------------------------------------------------------------
System S-810/20 S-820/80
Scalar processors 1 1
Scalar cycle time (ns) 14 8
Vector processors 1 1
Vector cycle time (ns) 7 4
Gates in logic 550/1,500 2,000/5,000
Switching time (ps) 350/450 200/250
Cache technology/chip density 1 Kb 6,900 Kb+2,500 gates
Cache access time (ns) 4.5 4.5
Cache size (KB) 256 256
VR access time (ns) 4.5 2.5
Vector registers (KB) 64 128
Floating point units 3 3 (Add&L+Add/Mult)
Pipelines per F.P. unit 2 4
Max Gflop 0.63 3 (2 if unchained ops)
Memory interleave 256
transfer rate (GB/s) 16
Main memory technology 16 Kb CMOS 64 Kb BiCMOS
Memory access time (ns) 40 20
Max. memory (MB) 256 512
Second-level memory 256 Kb DRAM 1 Mb DRAM
Access time (ns) 120
Max. memory (GB) 3 12
Transfer rate to main 1 GB/s 2 GB/s
I/O units 1 (32 ch.) 1 (64 ch.)
Total I/O throughput (MB/s) 96 288
Shipment date Dec. 83 Jan. 88
Appendix D CRAY supercomputers in Japan
=======================================
Table 1 Known CRAY supercomputers installed in Japan
(26 installed systems, sorted by installation name. 3 Y-MP2Es on order).
Customer name System Date Prefecture Sector
------------- ------ ---- ---------- ------
Aichi Inst. of Techn. X-MP/14se 88 03 Aichi Priv. Univ.
Asahi Chemical Y-MP2E/116 91 03 Shizuoka Chemistry
Century Research Corp. Cray-1 80 02 Tokyo Service Bureau
Century Research Corp. X-MP/18 88 01 Kanagawa Service Bureau
Daihatsu-Kogyo Y-MP2/216 90 07 Osaka Automobile
Honda R & D Y-MP8/364 90 09 Tochigi Automobile
Honda R & D X-MP/14 87 03 Tochigi Automobile
Isuzu Motor Y-MP2E/232 91 04 Kanagawa Automobile
Mazda X-MP/216 88 12 Hiroshima Automobile
Mazda Y-MP2E/232 91 02 Hiroshima Automobile
MITI/AIST X-MP/216 88 02 Ibaraki Gov.& Nat.Lab
Mitsubishi Elec. Lab Y-MP4/132 89 10 Osaka Conglomerate
Mitsubishi H.I. X-MP/116 90 05 Hyogo Heavy Industry
Mitsubishi Motor Corp. Y-MP4/116 89 10 Aichi Automobile
Mitsubishi Res Inst Y-MP2/116 89 11 Tokyo Research C.
Mitsubishi Research Inst. Cray-1 80 07 Tokyo Service Bureau
Nippon Telephone & T. X-MP/22 84 08 Tokyo Conglomerate
Nippon Telephone & T. Cray-2/4 87 12 Tokyo Conglomerate
Nissan X-MP/12 86 05 Kanagawa Automobile
Nissan X-MPEA/432 88 10 Kanagawa Automobile
Nissan Y-MP8/664 90 08 Kanagawa Automobile
Recruit X-MP/216 86 12 Kanagawa Service Bureau
Recruit X-MP/18 88 02 Osaka Service Centre
Sumitomo Chemical X-MP/116se 89 09 Osaka Chemistry
Tohoku University Y-MP8/4128 90 12 Miyagi Nat.University
Toshiba X-MP/22 85 02 Kanagawa Conglomerate
Toshiba Y-MP8/232 90 03 Kanagawa Conglomerate
Toyota X-MP/116 88 08 Aichi Automobile
Toyota Y-MP8/232 90 03 Aichi Automobile
Appendix E ETA supercomputers in Japan
======================================
Table 1 ETA supercomputers installed in Japan
(2 systems, sorted by installation name).
Customer name System Date Prefecture Sector
------------- ------ ---- ---------- ------
Meiji University ETA10-P 89 04 Kanagawa Priv. Univ.
Tokyo Inst. of Techn. ETA10-E8 88 05 Tokyo Nat.Univ.
Appendix F Fujitsu supercomputers in Japan
==========================================
Table 1 Known Fujitsu systems installed in Japan (59 out of 63 systems,
sorted by installation name). Internal Fujitsu systems not included.
Customer name System Date Prefecture Sector
------------- ------ ---- ---------- ------
Advantest VP-50 85 11 Tokyo Conglomerate
Air Force VP-50 87 08 Tokyo Gov.& Nat Lab
Asahi Kogaku PENTAX VP-30E 88 10 Tokyo Optical
Chiyoda Info. Service VP-50 86 04 Tokyo Chemistry ?
Chuo University VP-30E 87 10 Tokyo Priv. Univ.
Computer Techn. Integ. VP-2400/20 90 08 Service Bureau
Daikin Air Conditioner VP-100 87 03 Osaka Mechanical
Diesel Kiki VP-30E 89 02 Tokyo Automobile
Electric Power Lab. VP-50E 87 09 Tokyo Gov.& Nat Lab
Fuji Electric VP-50 85 12 Kanagawa Conglomerate
Fuji Electro-Chemical VP-50E 88 11 Tokyo Conglomerate
Hazama-gumi VP-30E 88 10 Tokyo Construction
ICFD (Fluid Dynamics) VP-200 86 04 Tokyo Research C.
ICFD (Fluid Dynamics) VP-400E 89 03 Tokyo Research C.
Inst.Nuclear Fusion VP-200 83 12 Ibaraki Gov.& Nat.Lab
Inst. Nuclear Fusion VP-200E 88 03 Ibaraki Gov.& Nat.Lab.
Inst.Space Aeronautic S. VP-200E 88 04 Tokyo Gov.& Nat.Lab
Ishikawajima-Harima VP-50 86 05 Kanagawa Heavy Industry
Jaeri (Atomic Energy) VP-2600/10 90 04 Ibaraki Gov.& Nat.Lab
Jaeri (Atomic Energy) VP-2600/10 90 04 Ibaraki Gov.& Nat.Lab
Kanagawa University VP-30E 87 08 Kanagawa Priv. Univ.
Kansai University VP-50E 88 08 Osaka Priv. Univ.
Kawasaki Steel VP-50 86 01 Chiba Metal
Keio University VP-50E 89 08 Kanagawa Priv.Univ.
KHI VP-50 87 06 Kawasaki Mechanical
Kobe Steel VP-200 87 06 Hyogo Metal
Kodak Japan VP-50E 88 11 Tokyo Chemistry
Kyoto University VP-400E 87 08 Kyoto Nat. Univ.
Kyoto University VP-2600/10 90 09 Kyoto Nat. Univ.
Kyushu University VP-200 87 08 Fukuoka Nat. Univ.
Matsushita VP-30E 87 08 Osaka Conglomerate
Matsushita VP-100 85 12 Osaka Conglomerate
Mitsubishi Kasei VP-50 86 07 Kanagawa Chemistry
Nagoya University VP-200E 88 03 Aichi Nat. Univ.
NAL (Space Techn.) VP-400E 86 12 Tokyo Gov.& Nat.Lab
NAL (Space Techn.) VP-2600/10 90 10 Tokyo Gov.& Nat.Lab
Nat. Astro. Observatory VP-200E 89 11 Tokyo Gov.& Nat.Lab
Nihon University VP-30E 87 12 Chiba Priv. Univ.
Nikko Shoken VP-2200/10 90 12 Financial
Nippon Kokan (NKK) VP-50 87 08 Kawasaki Metal
Nippon University VP-30E 87 12 Chiba Priv. Univ.
Nippon University VP-30E 87 12 Chiba Priv. Univ.
NTT VP-50 86 05 Kanagawa Conglomerate
Olympus VP-50 86 05 Tokyo Mechanical
Osaka Inst. of Techn. VP-30E 88 12 Osaka Priv. Univ.
Pacific Consulting VP-30E 89 01 Tokyo Consulting
Recruit VP-200 86 06 Tokyo Service Bureau
Recruit VP-400 86 06 Tokyo Service Bureau
Sharp VP-50 86 04 Osaka Conglomerate
Shibaura Inst of Techn. VP-30E 87 10 Tokyo Priv.Univ.
Shimizu VP-50 86 06 Tokyo Construction
Shionogi VP-30 87 05 Osaka Chemistry
SONY VP-2200/10 90 11 Kanagawa Electronics
Suukeikaku VP-30E 88 10 Tokyo Math. program.
Tokyo Electronics Univ. VP-100E 89 10 Tokyo Priv.Univ.
Tokyo University VP-100 86 11 Tokyo Nat. Univ.
Toray VP-30 87 08 Tokyo Chemistry
Toyota VP-100 85 08 Aichi Automobile
Toyota VP-100E 88 04 Aichi Automobile
Appendix G Hitachi supercomputers in Japan
==========================================
Table 1 Known Hitachi systems installed in Japan
(18 systems, sorted by installation name).
Internal Hitachi systems not included.
Customer name System Date Prefecture Sector
------------- ------ ---- ---------- ------
Bridgestone S-810/5 87 05 Tokyo Chemistry
CANON S-820/60 89 10 Kanagawa Conglomerate
Dainippon Print S-810/5 88 02 Tokyo Conglomerate
Hokkaido University S-820/80 89 02 Hokkaido Nat. Univ.
ICFD (Fluid Dynamics) S-820/80 88 10 Tokyo Centre
IMS (Molecular Science) S-820/80 88 01 Ibaraki Gov.& Nat.Lab.
JIP S-810/5 87 05 Chiba Centre
KEK (High Energy Lab) S-820/80 89 03 Ibaraki Gov.& Nat.Lab.
Metrology Agency S-810/20 87 11 Tokyo Gov.& Nat.Lab
MRI (Meteorology) S-810/10 85 11 Ibaraki Gov.& Nat.Lab
NDK Nippon El.Comp. S-810/10 87 02 Centre
Nihon University S-820/40 89 06 Chiba Priv. Univ.
Nissan Diesel S-810/5 87 04 Saitama Automobile
Olubis S-810/5 88 02 Shizuoka Centre
Suzuki Motors S-820/60 88 12 Shizuoka Automobile
Tokyo University S-820/80 88 01 Tokyo Nat. Univ.
Toyo Gum S-810/5 87 10 Tokyo Chemistry
Yamaichi Shoken S-820/60 89 04 Tokyo Finance
Appendix H NEC supercomputers in Japan
======================================
Table 1 Known NEC systems installed in Japan
(18 installed systems, sorted by installation name).
Internal NEC systems not included.
Customer name System Date Prefecture Sector
------------- ------ ---- ---------- ------
Aoyama University SX-1EA 88 10 Tokyo Priv. Univ.
Computer Engineer Ctr. SX-1A 88 12 Service Bureau
Daiwa Shoken SX-1A 89 08 Tokyo Financial
ICFD (Fluid Dynamics) SX-2 87 05 Tokyo Research C.
Japan Dev. Construction SX-JA 90 03 Construction
Japan Railway SX-JA 88 11 Tokyo Gov.& Nat.Lab
Kumagai SX-1 89 08 Tokyo Construction
Mazda SX-2A 89 09 Hiroshima Automobile
Obayashi Corp. SX-1EA 88 06 Tokyo Construction
Okayama University SX-1E 87 05 Okayama Nat. Univ.
Osaka University SX-2 88 01 Osaka Nat. Univ.
Port & Harbor Research SX-1E 87 12 Kanagawa Gov.& Nat.Lab
Recruit SX-2A 88 10 Tokyo Service Bureau
Sumitomo Metal SX-2 88 03 Osaka Metal Industry
Tohoku University SX-1 86 03 Miyagi Nat. Univ.
Tohoku University SX-2A 88 12 Miyagi Nat. Univ.
Tokai University SX-1E 86 09 Kanagawa Priv. Univ.
Tokai University SX-1 89 09 Kanagawa Priv. Univ.
---------------------END OF REPORT----------------------------------------