We managed to install easily the two cards a GTX470 and a GTX480 donated by ftpd (thanks).

ACEMD does not run on it as we know from previous tests on GPUGRID. We are now testing why.

It is strange because the app crashes with error 700 only on the Fermi card, but not on the other cards of the same machine.

gdf

Now ACEMD works on Fermi. Working on the optimizations.
Most likely on Monday we will have a working Fermi application out for Linux and possibly Windows.

gdf

So far, the performance of a GTX480 is only 30% faster than a GTX 285.
Fermi is completely different from a G200 chip, it's not just plug&play for a code optimized on the G200 as ACEMD.

We are trying to understand what's the problem. A factor two (100% faster) should be really the minimum for the hardware.

I'll keep you posted.
gdf

CC1.3 optimization did not happen overnight!

I take it that you are comparing a non-optimized Fermi to the optimized GTX285 ???

I think you said CC1.3 cards are now optimized to 200% from the original app via several updates. If so then compared to a non-optimized GTX285 would mean Fermi is about 2.6 times as fast, in a non-optimized to non-optimized comparison.

So which is it, an optimized or non-optimized comparison?

Either way, 30% better than a GTX285 is 130% better than last Friday!



PS. For anyone that doesn't know,
Fermi means 10 to the power of minus 15 metres (10^-15m).

Fermi means 10 to the power of minus 15 metres (10^-15m).

That's fempto not fermi...

In English, Fermi = femto

I know it means shut too. PS. There is no p in femto!

Did you compare the Fermi cards performance to that of an optimized app running on a GTX285 or a non-optimized app (the original ACEMD)?

Thanks,

The optimized application.

gdf

Then there is hope!
Thanks,

One question, if you succeed in optimizing your app to make Fermi cards run twice as fast as the GTX285, would that mean that someone crazy enough to buy a Tesla & crunch for GPUGRID.net would be 4 times faster than a Fermi & if so & the Fermi is only crippled through the Nvidia Driver, would that "potentially" mean that if Nvidia uncripples or "someone" hacks it, that the GTX480 would run 8 times faster than a GTX285 if you can get it to run twice as fast with a crippled GTX480?

---

One question, if you succeed in optimizing your app to make Fermi cards run twice as fast as the GTX285, would that mean that someone crazy enough to buy a Tesla & crunch for GPUGRID.net would be 4 times faster than a Fermi

DP as I understand it is crippled from 1/2 to 1/5th the SP speed, so for DP it could conceivably be a max of 2.5 times faster. However AFAIK GPUGRID doesn't use DP so there probably wouldn't be a speed increase at all. MilkyWay speed may increase to a reasonable level when they get the v3.x app working, but from early reports still slow compared to ATI.

So that's to say that the Fermi GTX480 was crippled from the start 1/2 & couldn't undergo a softmod, in the same way that 8xxx & 9xxx could become a Quadro. But still, if GPUGRID.net doesn't use DP it doesn't really matter...

---

So that's to say that the Fermi GTX480 was crippled from the start 1/2 & couldn't undergo a softmod,

DP speed is always slower than SP speed so the 1/2 speed is not because it's crippled. Dropping it to 1/5th is crippling it.

Current performance on the GTX480 is 35% faster than the fastest version of ACEMD on a GTX285.

There is little more we can do I think. The cuda3 compiler seems much worse at optimizing the code than the 2.2 version. If this is the case, then 3.1 could bring some improvement.

Tomorrow, we clean up and submit the new application in beta.

gdf
PS: no doubts this is much less than expected

In English, Fermi = femto

LOL! The first time I see this. It looks totally ridiculous.. but apparently is true.

MrS

---

Scanning for our furry friends since Jan 2002

Tomorrow, we clean up and submit the new application in beta.

Hooray !!!
I am all ready and setup for testing. I have a day job (east coast usa) so may not be able to provide up to the minute coverage but I will report any results as soon as I see them on my machine.

---

Thanks - Steve

We have now a good performance model of the application. The maximum we should have expected from Fermi is a reduction in time of 60% compared to a GTX275. Actually, we now think that this will be achievable in the future. The problem is that CUDA 3 is slower even on GTX200 by almost 15%. Indeed, we are 60% faster compared to both applications compiled for CUDA3 but compared to CUDA2.2 only 35%.

On our machine the running temperature is 91 degrees. Equivalent to a GTX275.
The only real problem is the price. Later I will report the time of the GTX470 and running temperature.

gdf

Boinc Messages:
NVIDIA GPU 0: GeForce GT 240 (driver version 19621, CUDA version 3000, compute capability 1.2, 512MB, 307 GFLOPS peak)

Boinc Tasks:
ACEMD – GPU molecular dynamics 6.03 (cuda) p45-IBUCH_101b_pYEEI...

Would this task run faster if I had CUDA version 2.2 or does it just depend on what you compile it on (2.2)?

Thanks,

Yes, it's what it's compiled with.

---

Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

.........The only real problem is the price.......
gdf

Thats a big problem - especially baring in mind that a 295 - taking up the same card space in a machine - takes out a 480, why should the world at large pay that premium for a hot & noisy 480, I think NVidia have the pricing wrong given the performance level.

Bad enough trying to find a 480 supplier, the other problem at present is no one with a 295 wants to let it go for a Fermi as they get no performance boost, and 295's are now EOL.

There is no way the market in general will be pushed around like that - it will just go ATI, they are pushing once loyal consumers over the edge.

Regards
Zy

WHEN Fermi IS Optimized by GPUGrid it WILL OutPerform the GTX 295!
Presently the highly optimized GTX295 just about outperforms a totally non-optimized Fermi, which is by the way not yet the finished article.
There will be a CUDA 3000 & driver update.
There will be a GPUGrid refinement for Fermi,
and there will be other Fermi's.

By the way, there is little price difference between the GTX480 and the GTX 295 and both are as hard to find.
That said, NVidia’s Internal Implosion of a strategy has recently been completely and utterly bonkers!

PS. Your RAC is still at Zero and yet you still keep banging on. Have you no shame?