ATI Catalyst 9.2 - finally Multi-GPU support!!!

[brentpresley]

ATI says that Cat9.2 now allows for Multi-GPU folding.


Anyone w/ multiple ATI video cards care to try this out and get back to us?

I'll be happy to help anyone out w/ the setup.

[huddy]

will do later today Brent!

[iDENTITi]

Yeah I can set up dual cards in my pc and try this out. I'll do it in a few hours.

[ColonelCain]

I wonder if it improves client stability for even one client... at its current state, it is unusable IMO.

[ColonelCain]

Read this over at the F@H forum...

If you want to utilize the new drivers, some speacial steps need to be taken (no need to fear!).

Folding Forum • View topic - ATI Catalyst 9.2 is out!

You must do a clean uninstall of your previous ATI drivers. Download 9.2, run it, but instead of selecting "Install", select the second one "Uninstall". You will need to reboot after finishing, then when it finishes, re-open 9.2, and install. Reboot again, and then, the only tricky part to it.
You can find the aticalcl.dll and aticalrt.dll files in your System32 folder, you must copy them into the folder that you have the GPU client installed in, and for the time being, rename the amdcalcl.dll and amdcalrt.dll (just put a 1 in the very front). Why the sudden name change? I have no clue.

In the link above, mhouston (big GPGPU guy with AMD), explains.

I also found in that subforum, mhouston suggests that very soon we will see a new client for ATI cards.

And in this thread: http://foldingforum.org/viewtopic.php?f=51&t=8245
mhouston suggests that either this next one, or the one after, will address the performance deficit b/t ATI and nvidia cards with the GPU2 client. He also mentioned a possibility of a 2.5x speed increase.

[brentpresley]

He also mentioned a possibility of a 2.5x speed increase.

If this is the case, you can bet I'll be dumping all my 9800GX2s in the near future.

[DaSickNinja]

So will this mean Stamford can quit gimpin the NV client?

[ColonelCain]

So will this mean Stamford can quit gimpin the NV client?

We can only hope...


Additionally, according to mhouston, their optimisation of the ATI GPU2 client is aimed at all product families, and then they will expand to optimizing for the 48xx series.

According to a blog that I was reading, the main cause of the CPU utilization by the ATI GPU2 client was the lack of local share cache for the SP arrays.

So, in other words, this is a limitation that is left from the R600 and RV670 generations.
The RV770 has local share cache, but ATI has been slow as molasses to expose the functionality to programers.

This is why nvidia kicks *****, is because ever since the G80, they have had local share cache.

AMD’s Folding performance explained, future development revealed Theo’s Bright Side Of IT

mhouston remarked that while this article jumped to many conclusions, they do recognize this as something holding back the 48xx series in folding, and Stanford may include access to local share cache in one of the next releases.

[Hawkeye]

He also mentioned a possibility of a 2.5x speed increase.

Well if that happens i'll have to finish my Quad Crossfire/ Phenom II build

[Violet_wind]

Hooray! I now have an even better reason to show off my HD 4830! (soon)

[ColonelCain]

Ok, went through and quoted the text that I found this in (I am not going to highlight much, as when people do you ONLY read that, and ignore everything else... the un-highlighted can change the significance):

The thread is titled: ATI Client limited to utilizing only 320 Shaders?!?!?!

Personally, ignore what I have quoted below, and just go to the link below, it is from this post on that is significant. mhouston is a big GPGPU guy working for AMD right now.
Folding Forum • View topic - ATI Client limited to utilizing only 320 Shaders?!?!?! [No]

Well, on the 4870, it depends a lot on the size of the protein. Basically, the force calculations are well behaved and can be unrolled to maximize processor utilization and we are using all of the 10 SIMD arrays (each with 16 5-wide VLIW ALUs), and our ALU pack rate is high. However, there are parts of the algorithm, like the update steps, that do not handle wide issue all that well, at least in their current form on ATI hardware, and may only make use of 1 or 2 SIMD arrays well. With large proteins, the force calculation dwarfs the other calculations, so the processor is used well. On small proteins, like lambda, the code is tuned up to the point now where the force kernels represent only a little more than half of the computation time, so the processor is not being used all that well.

However, the unrolling required to maximize usage during the force calculations causes overheads in other kernels, so there is a balance there and at least on the ATI core, we have some code that tries to make this tradeoff automatically, although it's only working well on the larger proteins.

In early cores, the processor was not well used for most proteins. We have gotten progressively better with each core update. To update toTOWs quote, lambda in the next release is using the processor much better now, but there is still some narrow issue points we are looking at as well as overall tuning of the algorithm. If a 48XX was used to it's full ability, you would expect a 2.5X performance increase clock for clock over a 38XX running the same code. However, there are some things that a 4XXX can potentially do that a 3XXX/2XXX board cannot that are also being looked at which may make the performance multiplier higher. Right now the concentration is getting the current code paths tuned all the way up.

We are trying to improve performance of ALL boards, i.e. the general code path. When that tops out, we can start looking at 4XXX and beyond specific features.

That is all that I will quote, or my post will go on forever. Just go and start reading at the post I linked on.

Read it. Good stuff.

Make's me happy that ATI users will be rewarded eventually, even if it is much later than when nvidia users were rewarded.

[ColonelCain]

OHHHHHH!!!!!

I have been reading some more, and according to mhouston, there will be a new core coming out soon, that will "reduce cpu usage on larger WU's dramatically"

Folding Forum • View topic - Frequent VPU Recover events with F@H 6.23 and Cat 9.1

[Harvestor]

hopefully that is true cain because my gpu client is really holding my 2nd smp back

[ColonelCain]

hopefully that is true cain because my gpu client is really holding my 2nd smp back

Stay tuned, I am going to post a new thread regarding my findings using a tweak that reportedly reduces CPU load.
Untill then, I would reccommend closing the second SMP, as the PPD of the 4850 is higher than the PPD of the 4850+SMP. So, run one SMP and the GPU2 client. But again, watch.

[Harvestor]

Ok i will, thanks for the tip.

[CapnBFG]

The thread is titled: ATI Client limited to utilizing only 320 Shaders?!?!?!

Personally, ignore what I have quoted below, and just go to the link below, it is from this post on that is significant. mhouston is a big GPGPU guy working for AMD right now.
Folding Forum • View topic - ATI Client limited to utilizing only 320 Shaders?!?!?! [No]

Make's me happy that ATI users will be rewarded eventually, even if it is much later than when nvidia users were rewarded.

I had noticed with my 4870 that one GPU2 instance loads the card less than 50% even though the fahcore process doesn't max out the respective CPU core. That's why I've been running dual GPU2 clients on my 4870 all along, which keeps the GPU utilization around 90%. I can add a third instance to max it out, but I've found that my PPD is better if I instead load the rest of my CPU up with dual SMPs.

I look forward to any improvements they make, although I won't hold my breath for anything incredible.

I also want to point out that the 2.5x improvement figure is a misstatement (and a very typical one at that). In the thread you linked with that figure, it was stated correctly in the OP and then misstated elsewhere. They state in the OP that for a 4870, current PPD is 5000, and best case is 10000 to 12500 (12.5/5=2.5, or going by shaders, 800/320=2.5). That means the new performance would be 2 to 2.5 times as much as current performance, and is therefore only 1 to 1.5 times more than current performance. So a maximum of 1.5x increase according to the numbers they gave, not a 2.5x increase. Sorry to nitpick, but incorrect usage of 'as much' versus 'more than' when stating figures is possibly my biggest pet peeve.

[ColonelCain]

Oh, that isn't nit picking at all, I am glad that you pointed it out.

That is odd that it doesn't use the full GPU. Have you tried starting up just one instance of GPU2, and let it run for a bit w/out any SMP clients on?

[logain]

When I was folding on my 3870 I had the same problem. Had to run 2 intances on one card to fully utilize it. I do not know about the first GPU client but GPU2 and ATI cards have had this problem from the beginning.

[huddy]

Hooray! I now have an even better reason to show off my HD 4830! (soon)

Violet getting a 4830? Nice

[CapnBFG]

Oh, that isn't nit picking at all, I am glad that you pointed it out.

That is odd that it doesn't use the full GPU. Have you tried starting up just one instance of GPU2, and let it run for a bit w/out any SMP clients on?

Yes, that was what I did when I first got the card. I let it run for about a day, and the the points generated were hardly more than my 3870 (except that with my 3870 a single instance showed higher GPU usage). Then I added on a second GPU2 instance to load it down, which almost doubled the PPD. Then since I had two CPU cores idling I decided to add SMP units on. I used Brent's info on the SMP affinity locker to get two SMPs running. After playing around with various configurations of just SMP, just GPU2, and both, I settled for two of each since that seemed to give the best PPD.