Why not integrated GPUs?

[VA1DER]

I'm curious, since the FAH benchmark program supports Intel integrated GPUs quite handily, why doesn't the actual client? My integrated GPU doesn't bench nearly as fast as my real one, but it benches about 4 times faster than my CPU. The benchmark is using OpenCL core 21, which is the same core as the client, so this suggests the core supports integrated GPUs. Any thoughts?

[MeeLee]

This topic has been beaten to death.
It appears that even with an Intel IGP, you're only folding at the speed of the slowest dedicated GPUs of now to 5 years ago.

It's indeed faster than the CPU, but CPU folding I guess, is easier to code than GPU folding, as CPUs are more flexible.. (not confirmed).
They want to reserve GPU folding for the fastest consumer GPUs on the market.

That, and the limited resources in making Intel IGPs to work with FAH.

[X-Wing]

There are a bunch of other threads on this topic, but basically it is too slow to use as a GPU slot. Even to augment the CPU calculations as like an accelerator, the speed up for the most common Intel iGPUs is only about ~20%, and as such doesn’t warrant the not inconsiderable effort to modify the client. The new graphics architecture coming out with Tiger Lake CPUs however promises to be much more powerful, and could, in theory, warrant that development effort.

[bruce]

As suggested above, GPU assignments have a limited range of deadlines. Some Projects are simple and others are orders of magnitude more complicated. WUs from projects that can be folded in a day or two on a CPU are unsuited for GPUs and, similarly, projects which can be folded in a few hours on a fast GPU may take many days on a slow GPU. The collection of all active projects does spread over a very wide range but projects start and finish without coordination with other projects, so the mix is dynamic. When there's no project that's perfectly suited for your hardware, FAH will assign a project which is not as well matched with your hardware.

Extending the range of supported GPUs to include slower GPUs would, in fact, slow down the average completion time of that batch of projects.

Since the same project might be assigned to the fastest GPU available and potentially the slowest GPU available (and that's a VERY wide rage of performances) projects are categorized by what they can be assigned to. It's fairly easy to manage the range of projects suited for CPUs and keep the assignment logic intact since CPU and GPU projects must be assigned from two independent batches of project and there's no overlap.

I have read that the latest GROMACS code does upport the intel iGPU and a portion of the calculation stream assigned to the CPUs can be off-loaded to an intel iGPU. This version of GROMAACS has not been ported to a FAHCore, and we have no proposed plan to upgrade to it. This is ABSOLUTELY not a promise of anything for FAH ... just a remote possibility for a development of a future FAH enhancement. I am intrigued by the possibility of speeding up the CPU assignments rather than slowing down the GPU assignments.

This is ABSOLUTELY not a promise of anything ... just a remote possibility for a development for a FAH future.

[bruce]

Another thing: iGPUs are typically available in laptops and similar hardware rather than in desktop machines that may be designed by gamers. In both cases, the heat generated by a GPU continuously processing a heavy compute load must be dissipated. Laptops tend to be intentionally power limited by not having a good way to get rid of the heat that FAH would generate on a GPU.

[JimboPalmer]

bruce mentions that heat is an issue for laptops in general, but iGPUs are on the same chip as the CPU.

Most GPUs are on a separate card with its own cooling system. If your case is robust, the CPU and GPU can both run full blast without effecting each other much.

With an iGPU, all the heat is controlled by the same cooling system, it is very possible the GPU will throttle the CPU and the CPU will throttle the GPU. Even in a big case, it is still the same chip/heatsink/fan.

So even with a AMD iGPU on a AMD CPU or a AMD GPU on a Intel CPU, you may find folding to be very constrained.

As always, let us know what you succeed at, we want to recommend successes.