Folding Forum

https://www.msn.com/en-us/news/technolo ... 0b51be3642

What is left except for the advent of quantum computers? And, what comes after quantum computers reach some limit?

A nearer goal could be AVX-512. https://en.wikipedia.org/wiki/AVX-512
F@H does not currently use AVX-512; there are 3 reasons:
Intel has implemented different subsets in each CPU generation
Intel has limited AVX-512 only in server CPUs, this is not Folding@Datacenter
Intel slows the entire CPU when executing AVX-512 instructions, so F@H would slow foreground tasks if it used AVX-512

There are rumors that AMD will implement a version of AVX-512 in all the Zen 4 CPUs, so home CPUs may finally have double the power when folding, if F@H writes a hypothetical Core_A9.

JimboPalmer wrote: ↑Tue May 31, 2022 2:13 pm F@H does not currently use AVX-512; there are 3 reasons:

Fourth reason - during initial testing of the first CPU folding core to support AVX there was little improvement in processing speed over AVX-256. It would have required creation and maintenance of an additional variant of the folding core for little gain to the F@h project. The later implementations of AVX-512 might have changed that cost-benefit calculation, but I don't know if that has been tested. Currently there is support for SSE2, AVX and AVX2.

JimboPalmer wrote: ↑Tue May 31, 2022 2:13 pm so home CPUs may finally have double the power when folding, if F@H writes a hypothetical Core_A9.

A9 is not hypothetical but already being worked on. It will offer support for GPUs.

Jonazz wrote: ↑Tue May 31, 2022 5:08 pm

JimboPalmer wrote: ↑Tue May 31, 2022 2:13 pm so home CPUs may finally have double the power when folding, if F@H writes a hypothetical Core_A9.

A9 is not hypothetical but already being worked on. It will offer support for GPUs.

Maybe there will be a version supporting GPUs, but that A9 test core failed to work out. The developer of that version is currently tied up with other obligations. There may be a release of a CPU only A9 core later this year, one of the research groups needs features that got support in a later version of Gromacs than used for the current A8 core.

A CPU centric core which needs GPU support feels like an April fools joke to me.

Is this where Bruce's thread asking about no GPU systems came from?
viewtopic.php?t=37165

aetch wrote: ↑Tue May 31, 2022 7:30 pm A CPU centric core which needs GPU support feels like an April fools joke to me.

Is this where Bruce's thread asking about no GPU systems came from?
viewtopic.php?t=37165

There is an option in the Gromacs code to use an attached GPU essentially as a vector processor to speed up some operations. An early test version of the A8 folding core did have that option enabled. It did work, and did speed up processing overall. The choice was made to not release the A8 core with that enabled as there were a number of support issues. The client would need updates to use and control the option. For example, the processing on the GPU did slow down processing by the regular GPU only folding core. There were also some other issues related to using the GPU enabled A8 folding core.

Bruce's thread mainly was looking for potential testers of folding on the Intel iGPUs using Core_22. It might have led to gathering enough data to determine which iGPUs were fast enough for regular GPU projects, if there was threshold processing power that was usable, and possibly might have led to using some of them as an adjunct to CPU folding through the Gromacs based folding core. AMD and nVidia cards potentially also can be used, as I recall the calculations were sent by the Gromacs code to the GPU using OpenCL commands.

Thanks
I couldn't see the sense in running a GPU accelerated CPU centric core running alongside a GPU dedicated core. I was envisaging a mountain of trouble with trying to run it on all the different combinations of system, up to my no-gpu system. The degradation of core21/22 performance does make sense though.

Bruce's thread, I did think it was to test the iGPU (I posted before realising this). At the time I thought it was due to the growing numbers of the iGPU, the GPU acceleration comments in this thread made me question that.

Joe_H wrote: ↑Tue May 31, 2022 2:46 pm

JimboPalmer wrote: ↑Tue May 31, 2022 2:13 pm F@H does not currently use AVX-512; there are 3 reasons:

Fourth reason - during initial testing of the first CPU folding core to support AVX there was little improvement in processing speed over AVX-256. It would have required creation and maintenance of an additional variant of the folding core for little gain to the F@h project. The later implementations of AVX-512 might have changed that cost-benefit calculation, but I don't know if that has been tested. Currently there is support for SSE2, AVX and AVX2.

I think they're going the opposite route.
Avx 512 can generate very precise results, but in terms of folding (finding variations on a constant), AVX512 is extremely power hungry and slow..

A few scientific articles appeared that deep learning (specifically boinc, which is closely related to fah), might be superseded by ai.
Essentially 4 or 8 bit shaders, using an entirely different method of testing variations than 512.
It is hands down hundreds of times faster and more energy efficient to focus on the variations, and thus don't have so much data that are constants being reprocessed.

In essence, Folding is a language of quantum computers (with answers ranging from -x to +x, rather than a binary language.

The computers that should be doing this work is quantum computers. And I think it's entirely possible that a few hours of quantum computer work, could easily result in more success, than a few thousands to tens of thousands of computing hours using the current traditional model.

But until we fully understand quantum computing, or can even begin to make programs for it (scientists think only regular computers will be able to make or generate quantum computer programs that are actually effective programs, as for humans to make such a program is probably going to be too difficult), we will be stuck with our current model until we adopt ai.

I see more like AI being the goto platform, rather than avx512.

Joe_H wrote: ↑Tue May 31, 2022 5:32 pm

Jonazz wrote: ↑Tue May 31, 2022 5:08 pm

JimboPalmer wrote: ↑Tue May 31, 2022 2:13 pm so home CPUs may finally have double the power when folding, if F@H writes a hypothetical Core_A9.

A9 is not hypothetical but already being worked on. It will offer support for GPUs.

Maybe there will be a version supporting GPUs, but that A9 test core failed to work out. The developer of that version is currently tied up with other obligations. There may be a release of a CPU only A9 core later this year, one of the research groups needs features that got support in a later version of Gromacs than used for the current A8 core.

I didn't know that, too bad. Thanks for the info!