Let's start an ATI folding guide here!!!

Mt_Goat

After seeing the success mas0n is having in this post I am ready to see more ATI card folders do better. Things that would help out are;

• what drivers are working well?
• do you need to set affinities for SMP folders running multi ATI GPU2 set-ups?
• What kind of overclocks does it take to crank up the PPD?
• .....any other suggestions?

Any info to help out would be great as I know there are several people running the ATI cards, me included. And I think it could help breathe new life into the team! A newer NVidia folder guide would probably be in order too.

mas0n

I threw together a sheet comparing two HD4870 cards. One is clocked at the stock core clock of 750MHz. The second card is running the core at 815MHz. I compared all WUs for which I had data for both cards. The card running stock is a 1GB version, while the overclocked card is a 512MB version. However, this should make no difference. You'll see that the results are a pretty mixed bag. I'll summarize here and add my data collected regarding WU frequency over the last 48 hours:

WUs 5741-5747: 3434 & 3830 PPD ~50% of all WUs
WUs 4742-4744: 4766 & 5319 PPD ~40% of all WUs
WUs 5735-5739: 2786 & 3087 PPD ~10% of all WUs

So I figure I am averaging 3902 PPD & 4351 PPD. Not bad.

ATI cards have come a long way these last few months in terms of GPU2 performance, but there is still much to be desired. I like to think that the 4870 would be capable of averaging 5500+ PPD once optimized. CPU usage is 100% of one CPU core per client, so process management is a must. I have seen no difference in performance that could be attributed to drivers since the 8.10 package, so getting the correct/best driver is a non-issue.

In the next 24 hours I'll build a sheet for my 8800GT and 9800GT to compare. I could be very wrong, but I feel like they are also averaging ~4300PPD.

Leonardo

ATI has come a long way these last few months in terms of GPU2 performance

Not so much ATI, but Pande Group is responsible. I'm sure you know, but the improvement in ATI GPU production is due to work units being distributed that process well on ATI GPU architecture - high numbers of 'slow' shaders. The trend had been work units that favored Nvidia GPU architecture - fewer but very fast shaders. But all the same, it's good to see those with ATI cards enjoy increased production now.

mas0n

Indeed Leo, I did not mean to imply that AMD/ATI was responsible, but that the performance increase has been quite steady for these cards. There has certainly been a tremendous effort from PG to make this possible.

I've edited my post a bit to better reflect my intent.

Mt_Goat

So if I have a Quad core running SMP I would use something like Process Lasso to set affinities as follows? (I am not being stupid but just trying to flush out answers so all will know without exception.)

• For 1 ATI card I would assign cores 0, 1 and 2 to the SMP client and core 3 to the GPU client.
• For 2 ATI cards I would assign cores 0 and 1 to the SMP client and core 2 to GPU 0 and core 3 to GPU 2.

Also:

• Is it possible to run 2 GPU clients on one CPU core?
• Do you really loose very much performance when CPU core affinities are not set?
• Are there settings someone should set in the Catalyst Control Center like AA?

Thank you mas0n for your valuable contribution!

mas0n

Mt_Goat wrote:

So if I have a Quad core running SMP I would use something like Process Lasso to set affinities as follows?

• For 1 ATI card I would assign cores 0, 1 and 2 to the SMP client and core 3 to the GPU client.
• For 2 ATI cards I would assign cores 0 and 1 to the SMP client and core 2 to GPU 0 and core 3 to GPU 2.

Precisely. I think it's also worth pointing out that when setting up a dual GPU2 configuration one must set the appropriate flags: -gpu n where n designates which GPU to use, starting with the topmost card: 0

• Is it possible to run 2 GPU clients on one CPU core?

Yes. However I would not recommend doing so with the current ATI GPU2 client as performance will suffer. Further optimization of the client could change this. It has in the past. Conversely, the Nvidia GPU2 client uses very few CPU cycles; often less than 1-2% of a single CPU core.

• Do you really loose very much performance when CPU core affinities are not set?

If you are running multiple clients, absolutely yes and especially if running the SMP client. If only a single GPU2 client is run on a PC, setting process affinity is not necessary although could be helpful if the machine is frequently used for other tasks.

• Are there settings someone should set in the Catalyst Control Center like AA?

No. Current settings in CCC that would typically be used for graphics performance will only affect APIs like DirectX and OpenGL. GPU2 does not use these. The original GPU1 client used DirectX and returned primarily unusable data. The ATI GPU2 clients run on CAL and Nvidia clients run on CUDA. It would certainly be possible for future drivers to include configurable features for these APIs, but nothing to date.

Mt_Goat

Thank you mas0n,
I hope this helps others as much as it helped me. I managed to squeeze an extra 1,230 PPD out of my 2 4830's with your info.