Slashdot Mirror
Top 500 Fastest Computers
epaulson writes "The Top500 list has been released for the first half of 1999. The number one machine remains ASCI Red. The biggest Linux machine is cplant at 129, and Avalon is number 160. The list is a ranking of results from the LINPACK benchmark, which is a Linear Algebra code, so things like distributed.net and SETI@home don't count. "
Secondly; is Blue Mountain completely up to speed yet? I seem to remember reading that it was going to be the fastest (albiet not by much) when all the processors were finally added. I dunno, maybe I was just smoking something or reading SGI press releases....
----
Every year during my review, I just pray the words "slashdot.org" aren't mentioned.
Yeah, I have an account on the 13th fastest computer in the world. Nothing quite as fun as sitting down and allocating 512 Processors to work on the formation of a planetary system. The Unicos/MK OS on Cray's is a very interesting example of microkernel design that is well done. Of course, the optimizing compilers on that baby are even more astounding.
neutrino
History has the relation to truth that theology has to religion-i.e. none to speak of. - Lazarus Long
This list is accurate at what it is trying to measure. The list is for general-purpose supercomputers and so this eliminates a number of machines that are much faster for dedicated tasks. The NSA, for instance, has machines that are very good at the task of factoring. These machines, however, are not capable of performing general tasks, and thus are not candidates for this list. In addition, you may say that there are other general purpose supercomputers out there that are faster, but not on the list. This is doubtful, as the machines have to be made by someone with experience at the rather specialized task of making ultra-fast systems. All of the major producers of supercomputers recognize the importance of such benchmarks and thus work with people to do Linpack tests. The importance of these benchmarks is that the company that produces high results will get more sales. The only computers now left out of this list would be those produced in secret by governments or other such entities. These groups fall behind because they do not have the existing knowledge and resource base needed to produce such computers. The end result: This list is quite accurate with perhaps a few Ultra Secret machines left out.
BTW, I work on two of the machines on this list.
neutrino
History has the relation to truth that theology has to religion-i.e. none to speak of. - Lazarus Long
The most useful computer in the world is the one you use.
----
Every year during my review, I just pray the words "slashdot.org" aren't mentioned.
I just went and got the 1000x1000 double precision benchmark from netlib.org. I grabbed the lapack library and g77 from the debian website (Debian 2.1/slink versions)
:)
On my 400MHz K6-2, I get 16 Mflops without optimization, 20 with -O3. Not quite what was listed in the performance document, but that might have been with a hand-tuned library.
For comparison, my home machine (a 300 HHz K6-2) gets 13 Mflops unoptimized, 20 with. It's running Debian 2.2pre/potato which uses egcs, so the optimization is probably better. Both machines have 100 MHz fsb and 1 MB L2 cache.
There's a fun java version on the LINPACK benchmark as well. I get 1.4 Mflops.
According to IC's web site, they have one Fujitsu VX/1 vector supercomputer and one Fujuitsu AP3000 massively parallel server, which is effectively a nest of 48 independent UltraSPARC systems linked together Beowulf style.
But when you look up Fujitsu on the Top-500 database, it turns out that only the vector supercomputer (VPP) series make the list, and none of their AP-xxxx systems.
For the VPP series, the entry-level for the top 500 is a twelve processor system rather than IC's single vector processor.
On the other hand, the AP-3000 would have enough total throughput at 45.6 Gflops to get on the list at number 172. But my guess is that it can only achieve that for problems that split into relatively big independent chunks.
That might be OK say for servers and big CFD models, but I suspect that the LINPACK test suite needs a much more fine-grained parallelisation, and would be much harder hit by communication latencies between nodes.
That's just a guess: perhaps any real supercomputer experts out there could say whether this sounds right ?
Even though I believe they modified the souce of Linux to run on all them processors, it is one of the advantages of Linux.
I am awaiting a press release from Redmond.
--
A machine cannot be included on the list if the owners don't submit the LINPACK results for consideration.
Christopher A. Bohn
cb
Oooh! What does this button do!?
That I think it is pretty cool that you can get an account on the 129th (CPLANT) fastest computer, if your work is valid. I just think that is too frigg'n cool.
Linux O Muerte!
"The only computers now left out of this list would be those produced in secret by governments or other such entities." - not true. I've worked on a number of systems that would spec on the list and aren't there.. Why? A lot of major corporations don't like to tell competitors what they are using.
Okay, I was bored, so I went through and counted some stuff:
/. won't let me).
The numbers won't add up correctly because several of the machines were credited to two co-builders. Or I could have made a mistake.
Company: total, # out of the top 10, highest rank
(I tried to make this line up but apparently
SGI: 182/500, 7/10, #2
IBM: 118/500, 1/10, #8
Sun: 95/100, 0/10, #54
H/P: 39/100, 0/10, #150
Fujitsu: 23/500, 0/10, #26
NEC: 18/500, 0/10, #29
Hitachi: 12/500, 1/10, #4
Compaq: 5/500, 0/10, #49
Intel: 4/500, 1/10, #1
Self-made: 3/500, 0/10, #129
SNI: 2/500, 0/10, #66
Tsukuba: 1/500, 0/10, #18
Siemens: 1/500, 0/10, #355
This ranking above looks very different than the ranking of the top five computers. For example, Intel, who is #1, is basically a non-factor in the supercomputer market, with a mere three other computers on the list. H/P and Sun, which don't even make the top 50, seem to have the mid-level supercomputer market locked up, with 134 computers between them. SGI, however, is still the undisputed leader, from the high end (7/10) to the mid and low ends of ths list.
10 PRINT CHR$(205.5+RND(1)); : GOTO 10
Wouldn't it be better to build a multi register CPU similar to MMX but on a scale in the thousands?
Intresting idea, but it does have it's flaws. For one, designing a new CPU is _really_ expensive. And as you add more parallelism, it gets even more complicated and expensive (look at Merced). The market for such CPUs would be very small, maybe a few hudred per year. As you may have noticed, even supercomputers are made as cheap as possible these days (eg. beowolf).
Needless to say, there would also be many technical difficulties. Feeding thousands of registers would require a very wide memory arcitecture, a few thousand bits might be a good start. I sure wouldn't want be the engineer responsible for designing a mobo for those CPUs..
Few architectural problems also. SIMD can be used effectively only when there is one operation that is done to a big array of data. eg you have an array of 1024 bytes, and you want to increase the value of each byte by one. However, not all code is like this. You might want to inrease the value of the first elemnt by one, the second element by two and so on. MMX just became useless, there is no paralelism here. Now we have a CPU that is working at a fraction of its full potential: of the 2000 or so registers, only two are used. There is other stuff too, but I lazy so..
Okay, I was really bored, so I did more stats. This time by country.
USA: 292/500, 7/10, #1
Japan: 56/500, 1/10, #4
Germany: 47/500, 0/10, #15
UK: 29/500, 2/10, #7
France: 18/500, 0/10, #47
Canada: 8/500, 0/10, #29
Sweden: 7/500, 0/10, #71
Netherlands: 6/500, 0/10, #146
Switzerland: 6/500, 0/10, #339
Italy: 5/500, 0/10, #36
Australia: 5/500, 0/10, #102
Korea: 3/500, 0/10, #78
Denmark: 3/500, 0/10, #275
Belgium: 3/500, 0/10, #286
Spain: 3/500, 0/10, #314
Finland: 2/500, 0/10, #53
Norway: 2/500, 0/10, #193
Austria: 2/500, 0/10, #392
New Zealand: 1/500, 0/10, #64
Luxembourg: 1/500, 0/10, #247
Mexico: 1/500, 0/10, #436
Summary: United States 292 vs. Everybody Else 208.
In the top ten, it's United States 7 vs. Everybody Else 3.
If you compile the stats by the country in which the corporation that made the computer is based, American companies are responsible for over 400 of the top 500 supercomputers (just about everything except the Japanese stuff).
10 PRINT CHR$(205.5+RND(1)); : GOTO 10