The state of ATI GPU folding

[farmpuma]

Andreas G at Nordic Hardware wrote,

" Radeon HD 5800 still lacks proper Folding@Home support

Folding@home is an important initiative that takes advantage of the enormous unused resources available with our PCs to better understand our protein diseases. Graphics cards has become the latest weapon in the fight against cancer and Alzheimer's as GPUs are well suited for simulating protein foldings with folding@home. When AMD launched Radeon HD 5800 with an unmatched theoretic performance of 2.7 TFLOPS many hoped that the card would boost folding performance, but that's not quite the case.

First of all there is no real support for Radeon HD 5800 with today's GPU client of folding@home and even if this can be circumvented performance is anything but impressive. With around 3000 PPD Radeon HD 5870 is on par with the precursor Radeon HD 4870 and far behind the NVIDIA GPUs.

The problem is that folding@home is not yet available as OpenCL, which means that AMD's client is based on the Brook+, a standard that is not really supported anymore. NVIDIA's GPUs use the CUDA based client while the AMD users will have to wait for an OpenCL client to appear.

In the end it means that Radeon HD 5800 doesn't have anything to add to folding@home today, which is a real shame considering the great potential of the architecture. Hopefully this will be remedied soon with an updated GPU client.

More information on folding with Radeon HD 5870 at Bright Side of News."

Since Stanford needed the help of Nvidia to develop the current GPU2 client, I guess we will need to wait until ATI/AMD has the time to lend a developmental hand. ATI, you are missing out on a substantial potential market. And power folders tend to run multiple high end video cards.

[flip-mode]

Thank god it's great for gaming. When it comes to GPGPU, I think Nvidia is going to be king of the hill. They're putting 110% of their effort to extend the usefulness of the GPU beyond graphics. For AMD / ATI, there's not nearly as much motivation to do so.

[Airmantharp]

You could really just call this the state of ATI 'GPGPU Computing.' Not only is the support not there for it, but I don't think that DAMMIT are really focused on GPGPU right now, on the hardware or software side. They seem to be having a heyday going after Nvidia's share of the gaming market, and are doing so with smaller GPU's and thus potentially bigger margins. The thing is, we see Nvidia still building larger GPU's to do the same amount of work, and it seems that this is because Nvidia is focusing on the GPGPU sector while focusing on the gaming sector as well, and bigger silicon is the result. Therefore AMD's GPU's are slower at GPGPU applications, and AMD is reluctant to compete with Nvidia as a forerunner in that market, since they lack whatever the extra silicon 'mojo' is needed in order to be efficient at GPGPU applications, and I consider the talk from AMD about RV870 supporting GPGPU to just be lip service to appease anyone who would say that they don't care outright. They care, and realize that there's a market, but they also realize that it's Nvidia's market, and if they're to survive and increase their sales, they'll have to do it by beating Nvidia in the gaming sector, not by splitting resources to focus on GPGPU as well.

*TL;DR, read flip-mode's post above.

[tejas84]

You could really just call this the state of ATI 'GPGPU Computing.' Not only is the support not there for it, but I don't think that DAMMIT are really focused on GPGPU right now, on the hardware or software side. They seem to be having a heyday going after Nvidia's share of the gaming market, and are doing so with smaller GPU's and thus potentially bigger margins. The thing is, we see Nvidia still building larger GPU's to do the same amount of work, and it seems that this is because Nvidia is focusing on the GPGPU sector while focusing on the gaming sector as well, and bigger silicon is the result. Therefore AMD's GPU's are slower at GPGPU applications, and AMD is reluctant to compete with Nvidia as a forerunner in that market, since they lack whatever the extra silicon 'mojo' is needed in order to be efficient at GPGPU applications, and I consider the talk from AMD about RV870 supporting GPGPU to just be lip service to appease anyone who would say that they don't care outright. They care, and realize that there's a market, but they also realize that it's Nvidia's market, and if they're to survive and increase their sales, they'll have to do it by beating Nvidia in the gaming sector, not by splitting resources to focus on GPGPU as well.

*TL;DR, read flip-mode's post above.

That is the most eloquent and intelligent post I have ever read on TR. Period!

@Airmantharp

You have hit the nail on the head. AMD cannot compete with Nvidia in the GPGPU space so going after the gaming share is the best solution.

In truth both firms will be financially fine while pursuing their respective directions but I think PC gaming really does not need over 2 Teraflops of single issue power. Dual issue is far more useful for things other than just games and this is where NV will make a killing.

Tegra and Fermi GPGPU will make Nvidia incredibly rich, so rich that they can acquire AMD...

[PRIME1]

Not to pour gas on the fire but.... I can't wait to see the kind of numbers Fermi churns out. My GTX260 seems to be pulling around 8000ppd.

Either way 3000ppd is probably better than any CPU. So it's probably still worth folding on a 5870 if that's what ya got.

Also it may be awhile for a new ATI client as AMD is waffling between OpenCL and Directcompute.

[Flying Fox]

Andreas G at Nordic Hardware wrote,

I thought the link was posted somewhere here or in one of those frontpage comments already? You need to read the forums more frequently.

Thank god it's great for gaming. When it comes to GPGPU, I think Nvidia is going to be king of the hill. They're putting 110% of their effort to extend the usefulness of the GPU beyond graphics. For AMD / ATI, there's not nearly as much motivation to do so.

I wouldn't say there is little to no motivation. It's probably just the bottomline talking. Remember they actually announced the Stream initiative before Nvidia is even associated with GPGPU? They had plans, but looks like all the cost cutting shelved the grand plans away. I'm sure the architects and product managers at AMD saw GPGPU in the future as much as Nvidia's people.

ATI

Still can't get over that, eh?

They seem to be having a heyday going after Nvidia's share of the gaming market, and are doing so with smaller GPU's and thus potentially bigger margins. The thing is, we see Nvidia still building larger GPU's to do the same amount of work, and it seems that this is because Nvidia is focusing on the GPGPU sector while focusing on the gaming sector as well, and bigger silicon is the result. Therefore AMD's GPU's are slower at GPGPU applications, and AMD is reluctant to compete with Nvidia as a forerunner in that market, since they lack whatever the extra silicon 'mojo' is needed in order to be efficient at GPGPU applications, and I consider the talk from AMD about RV870 supporting GPGPU to just be lip service to appease anyone who would say that they don't care outright.

I have to disagree. There is no evidence that given the proper tuning and/or refactoring/rewriting of the code, the AMD GPU client will still be slower than Nvidia's. Without further proof I am not just going to accept that AMD GPU is worse in terms of GPGPU applications in general. Right now we only have Folding and we simply don't have enough information.

Also, a bigger chip does not necessarily mean slower GPGPU performance. Barring a few timing critical (at or near realtime) applications, the massively parallel nature of GPGPU applications basically means the more execution units the merrier. So whatever communication overhead that may be perceived because of more physical chips may not be that significant in affecting overall performance.

Another potential trend that we simply don't know if it will appear until more GPGPU apps are on the market is the kind of "more complex" instructions that the 4+1 EUs of AMD's GPUs may benefit from. In earlier generations, synthetic benchmarks have already shown that for those more complex shader programs AMD's architecture had the advantage. If the next generation of GPGPU apps turns out to rely more on those instructions, the tide may very well turn. At this point however, I am not sure but am keeping an open mind.

[danny e.]

ATI

Still can't get over that, eh?

I dont think there is really much to get over as long as AMD keeps using the ATI branding.
My HD 5850 card says "ATI" on it.

[Krogoth]

What's more funny is that Folding@Home GPU client still has not fully utilize R7xx-era GPUs to their full potential. They are only using 320 "shading units"* out of 800 in a full R770 chip. Nvidia units have less "shading units"*, but they are clocked much higher which is why G9x and GT2xx GPUs are absolute monsters at F@H.

*- AMD and Nvidia differ in their definition of a "shading unit".

I doubt the problem is hardware-related. It is tied to software development and Nvidia has already been more aggressive at pushing their GPUs to do more GPGPU-related functions. Fermi seems to be build around this philosophy.

[farmpuma]

Andreas G at Nordic Hardware wrote,

I thought the link was posted somewhere here or in one of those frontpage comments already? You need to read the forums more frequently.

The link is from a one line news item from about three weeks ago. I kept it in an open tab until yesterday when I finally had the time to start this discussion.

Thank god it's great for gaming. When it comes to GPGPU, I think Nvidia is going to be king of the hill. They're putting 110% of their effort to extend the usefulness of the GPU beyond graphics. For AMD / ATI, there's not nearly as much motivation to do so.

I wouldn't say there is little to no motivation. It's probably just the bottomline talking. Remember they actually announced the Stream initiative before Nvidia is even associated with GPGPU? They had plans, but looks like all the cost cutting shelved the grand plans away. I'm sure the architects and product managers at AMD saw GPGPU in the future as much as Nvidia's people.

Yeah, I thought I remembered ATI/AMD starting the GPGPU effort with Stream computing before Nvidia jumped in with CUDA.

I'm surprised ATI/AMD hasn't been more motivated to take advantage of the tax deduction and hands on experience benefits of working directly with the Pande Group to develop a better client. Perhaps there is more going on than we know. Or perhaps AMD/ATI is to consumed with CPU damage control to see other opportunities.

[Ragnar Dan]

I'm surprised ATI/AMD hasn't been more motivated to take advantage of the tax deduction and hands on experience benefits of working directly with the Pande Group to develop a better client. Perhaps there is more going on than we know. Or perhaps AMD/ATI is to consumed with CPU damage control to see other opportunities.

There's a tax deduction for that? I'm telling you, if that's so, Pande/Stanford ought to have figured out a way for all the CPU time donors to deduct for their costs by now, too.

But then I'm still surprised ATi isn't making Intel chipsets & motherboards for people who like the little guy.

[Flying Fox]

I'm surprised ATI/AMD hasn't been more motivated to take advantage of the tax deduction and hands on experience benefits of working directly with the Pande Group to develop a better client. Perhaps there is more going on than we know. Or perhaps AMD/ATI is to consumed with CPU damage control to see other opportunities.

There's a tax deduction for that? I'm telling you, if that's so, Pande/Stanford ought to have figured out a way for all the CPU time donors to deduct for their costs by now, too.

I think 'puma means that DAAMIT can expense the R&D for this stuff as in "contributing to Stanford".

But then I'm still surprised ATi isn't making Intel chipsets & motherboards for people who like the little guy.

They were at one point, but they had to stop after AMD bought them.

[Ragnar Dan]

I'm surprised ATI/AMD hasn't been more motivated to take advantage of the tax deduction and hands on experience benefits of working directly with the Pande Group to develop a better client. Perhaps there is more going on than we know. Or perhaps AMD/ATI is to consumed with CPU damage control to see other opportunities.

There's a tax deduction for that? I'm telling you, if that's so, Pande/Stanford ought to have figured out a way for all the CPU time donors to deduct for their costs by now, too.

I think 'puma means that DAAMIT can expense the R&D for this stuff as in "contributing to Stanford".

Yeah, I was just having fun.

But then I'm still surprised ATi isn't making Intel chipsets & motherboards for people who like the little guy.

They were at one point, but they had to stop after AMD bought them.

I know. They can scarcely afford to give up markets. Intel could do what AMD decided to do, hoping to harm their competitor (outside of regulatory results, anyway), but not the reverse. Of course, that assumes it was profitable.

[Airmantharp]

ATi's chipsets for Intel weren't all that bad, aside from ACHI and USB issues that have continued to plague them. If AMD can get these parts fixed, they'd make an even better business case- of course they have to keep the drivers up. Chipset issues are one reason people like to stick with Intel.

[Adamantine]

The only reason Ati cards perform poorly with FAH is because Pande/Stanford didn't write very good code. Milkyway@home and Collatz Conjecture projects didn't optimize for the 5870 and it doesn't have much problem with scaling almost perfectly while using the same brook+ base.

So it's probably a money/political issue, meaning Ati didn't throw enough money at them while nvidia did.