Folding Forum

The real benefit will probably be on systems that don't support (or have) aa dGPU. That includes most OS-X systems and Intel laptops. The fundamental issue with laptops is always going to be heat. What PPD/W will we see when running whatever Watts can be dissipated at the system's thermal limit?

Offloading some of the calcultions to the iGPU and doing the rest on the CPU threads will most likely overlap more GFLOPS that using all CPU threads. yielding a higher PPD ... unless you run into thermal limitations.

As mentioned by MeeLee, it can be tricky to convince your BIOS that you want to enable both a dGPU (if you have one) and an iGPU. I gave up on that option without succeeding, but I couldn't use it for a higher PPD because GROMACS 2019 wasn't out yet and FAH may or may not decide to support it. My only benefit (at the time) was to avoid screen lag when that proved to be a problem for me.

With a dGPU, what steps should I use to get my iGPU to display the desktop? Would that work on a dual-boot machine (Linux/Windows)?

I gave up on that a while ago.
With Nvidia drivers, they take over from the igp drivers.
On Linux, at best, I can see the boot text, up till where the x-session windows manager starts (GUI).
It freezes there, and the GPU in the main pcie port takes over.

I've read the procedures for AMD GPUs are reverse from Nvidia.
First install Intel's drivers, and OpenCL. Then AMD.
Either way the dgpu will drive the display.

One con on using both dedicated and integrated graphics, is that if you're having only 2 CPU cores, and 2 dgpus, the little the igpu would fold on, might be less than what your dgpu might lose (due to sharing a core with the igpu).

in windows you can install gpu driver for both igpu+dgpu, both intel+amd and intel+nvidia works, i cant confirm it right now as im using amd ryzen cpu which didnt comes with an igpu
well, instead of wild guessing on the forum, its better assembly a test project using out of shelf gromac code, just to see how it performs

MeeLee wrote:One con on using both dedicated and integrated graphics, is that if you're having only 2 CPU cores, and 2 dgpus, the little the igpu would fold on,ight be less than what your dgpu might lose (due to sharing a core with the igpu).

Intel iGPU acceleration for FAH CPU slot only make sense if you have more than one CPU core assigned for CPU slot.

Case 1: CPUs and iGPU available; no dGPUs.

Using the iGPU to accelerate the GROMACS core running on the CPUs (Currently FAHCore_a7)

Case 2: dGPU(s) present and folding (Currently OpenMM using FAHCore_21 and potentially, FAHCore_22 soon) using whatever CPU resources are needed to support them.
* 2a) Without CPU folding
If an iGPU can be used to offload desktop Video generation or rasterizing can be shifted to the CPU, that minimizes the contention for OpenMM resources

* 2b) With CPU folding
Like case 1, if the iGPU resources benefit FAHCore_a7 or like 2a, off-load desktop video generation, Good. Otherwise disable the iGPU.

Intel iGPU acceleration for FAH CPU slot only make sense if you have more than one CPU core assigned for CPU slot.

Not really. A separate CPU thread is not going to be allocated to the iGPU by OpenMM. It'll be supported directly by the GROMACS code and hopefully, the experts a gromacs.org will wisely(?) allocate the total processing load to whatever devices are under the control of the GROMACS code just like AVX or SSE. The donor, of course can override that choice and disable the iGPU if the Donor decides it was an unwise decision -- probably based on thermal considerations.

Igp folding can share CPU resources with CPU slots.
Like Bruce said, as long as they're not stealing it from a thread assigned to a dGPU.

Intel IGPs are roughly as fast as 2 to 6 CPU cores, but run between 10-25W, vs 35-65W for CPUs. The latest (10th gen) Intel igp performs somewhere in between a GT 1030, and a GTX 1050.

Chances are much faster Intel IGPs and even dGPUs are on the horizon, as Intel is wanting some of that AI/deep learning server revenue.