ATi PPD

[ChasR]

Tim,
A point to discuss in the DAB.

Is the recent improvement in ATI productivity due to increased folding efficiency or by increasing the value of ATi WUs? I ask, because I noticed that the nVidia WU p11175 is described as Ab42gb, 627 atoms, 1352 points while ATi WU p11292 is described as Ab42gb, 627 atoms, 2224 points. That 64% difference in value accounts for almost all of the ATi improvement, if they are the same WU. Before I get upset about it, It would be nice to know if it's true. If it is, what's next? Creating AMD specific WUs and valuing them to bring them closer to Intel production? I suppose ultimately science will prevail. Pande Group can not afford to artificially close the gap too much, lest they influence buying decisions in a way that negatively impacts the amount of science done.

 

[ihrsetrdr]

Will do Charlie. Do you mind if I quote you, verbatim?

 

[ChasR]

Fine with me.

Along the same lines, I note that several ATi donors are running two instances on their single gpu cards because they make more ppd. If they actually double it, fine they aren't hurting anything, at least until WUs are introduced that more fully saturate the shaders.

 

[ihrsetrdr]

Topic is posted, no replies yet. You should actually be the DAB rep rather than me, Charlie. I am just much less involved in beta testing than I used to be, more 'distracted' with RL matters as of late.

I am so way out of the loop with GPU folding, I can't conceive of running 2 clients simultaneously on a single core card.

 

[ihrsetrdr]

Some replies:

We are benchmarking as posted in the Points FAQ, i.e. on NVIDIA hardware, for comparison. We're not artificially closing any gaps -- the cards do what they can do. There's a natural reason to be direct here: artificially increasing points means that we're *decreasing* the science output by encouraging people to run on hardware they shouldn't. I feel pretty strongly that we have to keep the points in line with the scientific output and let donors decide from there what's best for them.

Then, too, you'll find people who find a conspiracy under every rock. FAH needs to maintain a vendor-neutral stance as much as possible (again remembering that the objective is to maximize science).

As software is initially developed, vendor-neutrality is not always possible due to support costs and various support issues, but it still must be FAH's goal. In the early days of the GPU, only ATI was supported (by way of Brook). Then a core version was developed for NV (by way of CUDA). Now a new core is available for ATI (via OpenCL) which is a couple of generations newer than the original. (The hardware is also several generations newer, but that's a separate issue.) If all three Cores ran on both platforms (not possible, but still worth thinking about) there would be PPD differences due to hardware differences and PPD differences due to software differences, but nowhere along the line would an artificial points adjustment be a good thing.