LESS PPD WITH AMD Radeon HD 7950 THAN AMD Radeon HD 6870

[P5-133XL]

Points issues aside, do you realize that PG does not make the cores. There is a separate open source organization that creates the GROMAC's cores. PG licenses their product and then designs the projects and the clients. I'm sure that there is significant feedback from PG but, to me, it is unfair to blame all core optimization issues on PG.

[mdk777]

[/quote]to me, it is unfair to blame all core optimization issues on PG.
[/quote]

So noted in my response:

It is what it is, explain the lack of resources, or the complexity and collaborative nature of the project, fine.

My point is directed toward answering questions directly and honestly.

If someone asks why a 3x power card performs so poorly; then just give them the straight answer.

If someone asks if they can run their AMD7970 to do MAYA?
Here is the answer I give:

The difference in hardware between the $3000 version of the professional card(FirePro W9000 6GB) and the $600 consumer card(SAPPHIRE Vapor-X 100351-6GVXSR Radeon HD 7970 GHz Edition 6GB) is very small.
However, writing optimized drivers and submitting them for qualification to all the major rendering,and professional CAD program companies is very expensive.
Consequently, the professional drivers will not run on your consumer card. Hence, you will see a significant (2x to 10x) reduction in performance running on what is essentially very, very similar hardware. When you buy the professional card, you are paying for the entire software support package.

It is what it is. Your hardware is great(amd7970), but if you want the best results, you are better off buying a significantly reduced hardware (AMD FirePro V5900 2GB, for around $450). While it is far inferior hardware to your AMD7970, the results you see in your profession programs will be far superior.


simple answer to a simple question.
Why doesn't my hardware perform in FOLDING?
The software is not yet optimized.

Explaining why is great, but blaming memory, bus size, etc. etc. etc. is disingenuous.

[cristipurdel]

For me it is kind of strange when fah requests the "best" hardware but does not fully utilize it.
Since HCC is up and running, my 6490M card which definitely cannot make the wall time here, still manages to crunch a WU in 30 minutes over to "the other side".
Anyway, I guess that in one month there will be a clear comparison just by looking at the PFLOPS for both projects.

[k1wi]

One thing to note about FAH and gromacs is that they spend a lot of time hand-optimizing the code so that it works incredibly efficiently on a given platform.

If you compare NVidia's development with AMD's development it really is no surprise that NVidia's is so far ahead...

Take the past three generations of AMD cards, you have three fundamentally different architectures: VLIW5, VLIW4 and most recently, a big shift with GCN. Big changes each step of the way. It has to be noted that for AMD, they actually got the utilisation of VLIW5 pretty high, apparently between 4 and 4.5 according to mhouston, which is a pretty good effort as that architecture was very challenging to hit anywhere near peak efficiency.

Compare that with NVidia: Up until the latest series, where there was a relatively big shift (ironically Nvidia shifted from fewer more powerful 'cores' to many weak 'cores' while AMD has done the reverse), the underlying architecture had evolved over recent generations, rather than radically differ. But while NVidia has had a large shift between the 500 series and 600 series, software and firmware wise it's been a much simpler path forward. If AMD's hardware architecture has been all over the place it's firmware/software evolution has been even worse (and its reputation historically on drivers are pretty poor, though from the forums it looks like mhouston has in the past put in a lot of effort into resolving driver issues). When results need to uphold such scientific integrity and rigor, development doesn't happen as fast as it does in Web 2.0 sites...

So while the latest gen of architecture is much different, getting a workable client out for Nvidia has been much easier. Any OpenCL core on the GPU is going to be far less mature than the CUDA implementation. Furthermore, to simply say "AMD outperforms NVidia in some other GPGPU applications" is incredibly disingenuous because so many factors relying on the actual calculations being undertaken alone.

Finally, it has been said for a while now that for a long time ATI folding has been 'waiting' on larger WUs to be processed, in order to take better advantage of all the stream processors. I personally don't think hardware should dictate the science - they shouldn't release large implicit solvent WUs just because they have the GPUs to do so. Although, on the CPU side of the ledger much larger explicit solvent proteins do exist (BigAdv anyone?), that is what I find makes this quote most interesting:

With recent advances in both cores and completion of our testing of these capabilities to ensure agreement, we are now confident we can do the same work on both cores.

In other words, there is the possibility that formerly CPU only, larger explicit proteins, may soon be able to be crunched on the GPU. That fact seems to have been overlooked somewhat in the broader analysis of that blog post. What the post means in terms of platforms supported, computational efficiency, availability and numerous other aspects is purely speculation until they roll out...

In other words, I am optimistic that the relative 'quiet' lately has hidden what is hopefully a large amount of development work behind the scenes. I for one am keeping my eyes on the developments ahead as hopefully all the different components are finally lining up in a manner that means GPU folding can really make a big leap forward.

[mdk777]

Well, just an update.

Today AMD released CAT 12.11 Beta.
CAT 12.9 Beta had a nice increase in OPEN CL score on LuxMark V2.0

AMD 7970 {Tahiti, GPU, 32,1050}open cl 1.2

yields a score of 1619

12.9 yielded 1800 and now CAT 12.11 yields 2089

Point: In this open CL test, drivers have improved efficiency by 30% in just the last few months.

Furthermore, to simply say "AMD outperforms NVidia in some other GPGPU applications" is incredibly disingenuous because so many factors relying on the actual calculations being undertaken alone.

These scores represent the raw processing "POTENTIAL" of the AMD card in open CL.

The point has always been to encourage PG, gromacs, (or whoever is responsible for the cores) to utilize that potential in a more timely fashion.

I recently read an interesting paper which discussed program optimizations for the 4870...showed up to 10X improvements in processing with the proper code usage and optimization...

Only problem is that is was published 3 years after the card was already obsolete from an efficiency and ROI standpoint.