New NVIDIA Tesla GPUs Reduce Cost Of Supercomputing By A Factor Of 10

[[I.R.A]_FBi]

This is proof there is a gt300. So where is our desktop cards huh nvidia?

As previously announced, the first Fermi-based consumer (GeForce) products are expected to be available first quarter 2010.

 

[jessicafae]

It looks like the single precision performance of C2070 (NV100 Fermi) is only 35% better than the previous generation C1060 Tesla (GT200 based). Granted for HPC the double precision is most important for this product. This will be a very interesting HPC/Supercomputer part. But gaming uses single-precision mostly, so the Geforce fermi will be interesting....

www.brightsideofnews.com/news/2009/11/17/nvidia-nv100-fermi-is-less-powerful-than-geforce-gtx-285

This table is coming from here.
This is a small comparison between three generations of Tesla parts:
3Q 2007: C870 1.5GB - $799 - 518 GFLOPS SP / No DP support
2Q 2008: C1060 4GB - $1499 - 933 GFLOPS / 78 GFLOPS DP
2Q 2010: C2050 3GB - $2499 - 1040 GFLOPS / 520 GFLOPS DP
3Q 2010: C2070 6GB - $3999 - 1260 GFLOPS / 630 GFLOPS DP

 

[Benetanegia]

It looks like the single precision performance of C2070 (NV100 Fermi) is only 35% better than the previous generation C1060 Tesla (GT200 based). Granted for HPC the double precision is most important for this product. This will be a very interesting HPC/Supercomputer part. But gaming uses single-precision mostly, so the Geforce fermi will be interesting....

www.brightsideofnews.com/news/2009/11/17/nvidia-nv100-fermi-is-less-powerful-than-geforce-gtx-285

Remember that the GT200 numbers are with dual-issue (MADD+MUL= 3 ops/s) and that Fermi is showing FMA numbers (2 ops/s). In reality GT200 could never or almost never have access to the extra MUL so it was actually only MADD (2 ops/s) most of the times and specially in games only MADD was in use. The actual number was in reality ~650 Gflops for the GT200 cards. Performance has mostly been doubled and if you consider what they say in update #2, the numbers shown for the Tesla's have the performance hit from ECC added to the equation. According to a document released by Nvidia some time ago ECC can hurt performance for as much as 20%, 5-20% they said, depending on the application. That's why GeForces will have ECC support disabled. They also say that Tesla cards are lowest clocked Fermi product in order to meet the high stability required for HPC qualification, they have to work for years of 24/7 operation. All things taken into account Fermi more than delivers a 2x increase in performance, at least on paper. We will find out in Q1 2010.

 

[jessicafae]

Remember that the GT200 numbers are with dual-issue (MADD+MUL= 3 ops/s) and that Fermi is showing FMA numbers (2 ops/s). In reality GT200 could never or almost never have access to the extra MUL so it was actually only MADD (2 ops/s) most of the times and specially in games only MADD was in use. The actual number was in reality ~650 Gflops for the GT200 cards. Performance has mostly been doubled and if you consider what they say in update #2, the numbers shown for the Tesla's have the performance hit from ECC added to the equation. According to a document released by Nvidia some time ago ECC can hurt performance for as much as 20%, 5-20% they said, depending on the application.

This is interesting. I am guessing that Nvidia had to adjust the official GFLop numbers for Tesla (not dual-issue SP) to bring them closer to reality because of the big HPC contracts they are negotiating.

The latest CPUs are really not that far behind Tesla these days for HPC :: Fujitsu's Venus SPARC64 VIIIfx can do 128Gflops double precision in around 40watts (compared to the new Tesla C2050/C2070 official 520-630 GFlops DP in 190watts). And IBM Power7 will be around 256Gflops per CPU when deployed in 2010/2011 for NCSA's "Blue Waters" supercomputer.

I did find the last statement of update#2 interesting.

from here. Tesla cGPUs differ from GeForce with activated transistors that significantly increase the sustained performance, rather than burst mode.