GIMPS HITS {insert number here} COMPUTERS

[jinydu]

The virtual machine's sustained throughput* is currently 20170 billion floating point operations per second (gigaflops), or 1675.6 CPU years (Pentium 90Mhz) computing time per day. For the testing of Mersenne numbers, this is equivalent to 720 Cray T916 supercomputers, or 360 of Cray's most powerful T932 supercomputers, at peak power. As such, PrimeNet ranks among the most powerful computers in the world. (*Measured in calibrated P5 90Mhz, 32.98 MFLOP units: 25658999 FPO / 0.778s using 256k FFT.)

Current PrimeNet Atomic Clock UTC Time is Saturday 22 October 2005, 09:25:13

Amazing! I think that last time I checked, it was at 16 or 17 teraflops.

This is the first time I've seen it past 20 teraflops.

[moo]

**Claps and dances around**

[Peter Nelson]

Impressive!



We can by all means celebrate this achievement

BUT

"The virtual machine's sustained throughput* is currently 20170 billion floating point operations per second (gigaflops), or 1675.6 CPU years (Pentium 90Mhz) computing time per day. For the testing of Mersenne numbers, this is equivalent to 720 Cray T916 supercomputers, or 360 of Cray's most powerful T932 supercomputers, at peak power. "

This statement is VERY MISLEADING because it implies (to those less knowledgeable about supercomputers) that T932 (around 60 GFlop on 32 processors slightly less about 1.7GF per processor) is Cray's fastest computer. It ISN'T! THIS STATEMENT SHOULD BE CHANGED!

Better to omit "MOST POWERFUL" and to compare with more recent eg.

T3E systems 6-2048 Alpha@450 or 32-2048 Alpha@600
These give 2.5 TERAFLOPS for a 1024 processor system
So theoretical peak for a fully loaded one is 5 TERAFLOPS.

So you could say we reached FOUR T3E systems.

BUT

We are FAR short of their top systems

eg.

CRAY X1E from 16 up to 8192 processors delivers UP TO 147 TERAFLOPS in a single system.

It will take us a while to catch up with that.

We currently achieved 14% of the REAL Cray. :

Incidentally if we wanted to compare with OLD Crays, the first (Cray 1, 1976) did "only" 133 Megaflops. We could feel even better if we compared Primenet to that (150 thousand times faster), but I hope you take my point that to do so (as a comparison with "a supercomputer") is very misleading to the general reader.

[Peter Nelson]

Just out of interest, 20 TFlops does put Primenet on a similar performance to some of the TOP 10 OF the top500 supercomputer list (from 6 months ago - the new November 2005 one comes out in about a week so most will get bumped down the league table by faster machines). We could probably be ranked somewhere from #7 to #12 in June's list if we qualified (but we didn't achieve our current performance back in June did we?)

Rank Site
Country/Year Computer / Processors
Manufacturer Rmax
Rpeak (figures in Gigaflops)

#1 DOE/NNSA/LLNL
United States/2005 BlueGene/L
eServer Blue Gene Solution / 65536
IBM 136800
183500

#2 IBM Thomas J. Watson Research Center
United States/2005 BGW
eServer Blue Gene Solution / 40960
IBM 91290
114688

#3 NASA/Ames Research Center/NAS
United States/2004 Columbia
SGI Altix 1.5 GHz, Voltaire Infiniband / 10160
SGI 51870
60960

#4 The Earth Simulator Center
Japan/2002 Earth-Simulator / 5120
NEC 35860
40960

#5 Barcelona Supercomputer Center
Spain/2005 MareNostrum
JS20 Cluster, PPC 970, 2.2 GHz, Myrinet / 4800
IBM 27910
42144

#6 ASTRON/University Groningen
Netherlands/2005 eServer Blue Gene Solution / 12288
IBM 27450
34406.4

#7 Lawrence Livermore National Laboratory
United States/2004 Thunder
Intel Itanium2 Tiger4 1.4GHz - Quadrics / 4096
California Digital Corporation 19940
22938

#8 Computational Biology Research Center, AIST
Japan/2005 Blue Protein
eServer Blue Gene Solution / 8192
IBM 18200
22937.6

#9 Ecole Polytechnique Federale de Lausanne
Switzerland/2005 eServer Blue Gene Solution / 8192
IBM 18200
22937.6

#10 Sandia National Laboratories
United States/2005 Red Storm, Cray XT3, 2.0 GHz / 5000
Cray Inc. 15250
20000

[jinydu]

Hmm... Do you think Primenet would ever reach 100 teraflops?

[ppo]

How is sustained throughput defined? I am afraid that it could be a little misleading, for instance I had to define twice my computer, since I am sometimes runnig two copies of Prime95, but that does not means that the sustained throughput is the double of my single cpu.
It would be nice to have some statistic based on the work really done, for instance using the mean number of exponent cleared in a day and the effort needed, based on their size, but that is asking too much, for so little information gained.
Keep the good work going on

[fatphil]

Originally Posted by ppo

How is sustained throughput defined? I am afraid that it could be a little misleading, for instance I had to define twice my computer, since I am sometimes runnig two copies of Prime95, but that does not means that the sustained throughput is the double of my single cpu.
It would be nice to have some statistic based on the work really done, for instance using the mean number of exponent cleared in a day and the effort needed, based on their size, but that is asking too much, for so little information gained.
Keep the good work going on

Sustained throughput means the throughput that can be sustained.
Unlike peak throughput, which is merely the peak throughput.
Simple, eh?

Like cheetahs. Amazing peak throughput - 100km/h, but they can't sustain it.
A horse can get further in a day than a cheetah.

[ppo]

how is it computed ?

[cheesehead]

Originally Posted by ppo

It would be nice to have some statistic based on the work really done, for instance using the mean number of exponent cleared in a day and the effort needed, based on their size,

That is how PrimeNet does calculate throughput, except that it's on results reported to PrimeNet (which I'm guessing is what you meant by "work really done") rather than exponents cleared. (E.g., completion of a factoring assignment that does not find a factor counts as throughput, though it doesn't clear any exponent.)

Then the "sustained" throughput is just the seven-day running average (same figure as the GFLOPS/s total under "Last 7 Days Average", under "------- Aggregate CPU Statistics, P90 Units* -------", under "Hourly World Test Status Summary").

[ppo]

I see, thanks, now I understand better.

[ppo]

How is the rate of results in error? is it improving with the years or not ?