Folding Forum

There are a lot of Gen7 iGPUs. If you restrict them to small projects, they should be good to go.

This is F@H's Position:

"GPUs must be Intel Gen 9. Support for Gen>9 can be expected in the future. For Gen<9, no support is planned."

Even with Gen 9, very few points are generated. F@H uses Integers as Points, no fractional points are possible.

(From Microsoft's perspective ... and probably Apple, too): Early iGPs were just strong enough to display the desktop. Nobody was predicting the demands of current games. When FAH was ported to GPUs, it was designed to use hardware that only existed on game-ready hardware ... namely NVidia and AMD GPUs. Years passed and hardware became stronger, with nV and AMD being driven by the continuous increases in the demands of newer games -- and by competition with each other. Meanwhile, the iGPUs from Intel and AMD were enhanced in parallel with their CPUs ... but that progress has been MUCH slower than the gaming world and the AI world.

Early OpenCL drivers from Intel/Microsoft were bug-ridden in areas required for stream computing while just meeting all the requirements of displaying a desktop. About Gen 9, the drivers finally supported the computational functionality of games and concurrently, of FAH. Hardware speed was still being driven mostly by desktop and YouTube / Netflix, not by games so it still MUCH weaker than AMD/NV GPUs. Nevertheless, FAH has managed to isolate some really tiny proteins that need to be folded which totally waste the capabilities of game-ready GPUs -- and they fit nicely between the FAH projects that can be done by traditional CPUs with 2,3,4...12 threads and game-ready GPUs. They fit well with the capabilities of IGPs Gen9 and above.

If that happens to be the hardware you have, good.

codysluder wrote:There are a lot of Gen7 iGPUs. If you restrict them to small projects, they should be good to go.

You won't say the same when you'll see how slow it already is on Gen 9 ...

Maybe iGPUs < gen 9 can still be used to accelerate FAH CPU cores in future by using offload?
http://manual.gromacs.org/documentation ... mance.html

Yes, maybe ... but current release considers them as regular GPUs and will use OpenMM core (Fahcore_22) ...

Intel <9th gen GPUs are rated at around 200Gflops (peak), while >9th gen are rated at >500Gflops, some as high as 2Tflops.
When offloading CPU 32bit fpp, allowing the CPU to focus on 64bit DPP operations, the question can be asked, if the GPU can also access L-cache data, or the CPU needs to write the Lcache into RAM first.
If this is the case, any GPU below 500Gflops isn't going to accelerate the work, but actually slow it down, and cost more power.
If Intel IGPs are able to directly draw data from L-cache, they could accelerate work quite well.
They certainly will accelerate Atom processors. But at 200Gflops, you're looking at ~1-3k PPD (while a dual core Atom/celeron processor running at ~1,66-1,85Ghz gets about 1k PPD as well).

iGPU offload for FAH CPU core could give upto 20% speed improvement. So if 4 core 4Ghz CPU does 100k PPD then with iGPU offloading could be max 120k PPD total.

foldy wrote:iGPU offload for FAH CPU core could give upto 20% speed improvement. So if 4 core 4Ghz CPU does 100k PPD then with iGPU offloading could be max 120k PPD total.

I'm not sure how the new core goes.
But the more cores the CPU has, the smaller the GPU acceleration will be in percent.
Intel mainly equipped their older CPUs with an IGP that's capable of playing back videos, show a GUI, and do basic tasks.

The Intel GMA (1080p video playback, gaming at DX9 800x600 max), Intel HD3000/4000 series (1440p video or 480p DX10/720p DX9), or intel UHD (with 4k video playback, <720p DX11/1080p DX9) gave different results.
Intel UHD being <500 Gflops (~20-30k PPD), HD 3000/4000 closer to 200Gflops max, and GMA barely 100Gflops.
It'll also depend on the mobile or desktop variant.
And even with these, depending on which model, you could end up with anywhere from 4 or 6 to ~24 or 32 cores/shaders; so performance would vary.

Would be interesting to do tests even on older CPUs (<7th gen).
They use the Beignet drivers for OpenCL.