CPU folding on iGPU?

[foldy]

Latest Gromacs 2019 now supports offload of tasks to CPU integrated iGPUs using OpenCL 1.2.

So maybe a next FahCore_a8 for CPU could use the Intel or AMD CPU integrated iGPU to help CPU folding?
As these iGPUs were mostly too slow for GPU folding they finally could be used to help CPU.

http://manual.gromacs.org/documentation ... mance.html

PME long-ranged interaction GPU offload now available with OpenCL

On supported devices from all supported vendors (AMD, Intel, NVIDIA), it is now possible to offload PME tasks to the GPU using OpenCL. This works in the same way as the former CUDA offload. A single GPU can now be used to accelerate the computation of the long-ranged PME interactions. This feature means that only 2-4 CPU cores per GPU will be about as fast as the 2018 version that needed many more CPU cores to balance the GPU. Performance on hardware that had good balance of GPU and CPU also shows minor improvements, and the capacity for hardware with strong GPUs to run effective simulations is now greatly improved.
Intel integrated GPUs are now supported for GPU offload with OpenCL

On Intel CPUs with integrated GPUs, it is now possible to offload nonbonded tasks to the GPU the same way as offload is done to other GPU architectures. This can have performance benefits, in particular on modern desktop and mobile Intel CPUs this offload can give up to 20% higher simulation performance.

[toTOW]

I think we never used Gromacs for GPU in FAH ... GPU cores are based on OpenMM.

[foldy]

I know the OpenMM based FAH GPU cores do not run on the CPU integrated iGPUs. But now the Gromacs 2019 allows a CPU core to use the CPU integrated iGPU. So the iGPU is seen similar to the CPU SSE or AVX extensions. New gromacs supports iGPU and FAH does not yet use iGPU - so bring it together

[bruce]

You might check and see if this is included in enhancement list for OpenMM when the short-range forces are being processed on the GPU.

[toTOW]

I've seen some BOINC projects trying to use Gromacs on CPU and GPU simultaneously, but it was never very efficient ...

[foldy]

Gromacs says upto 20% speedup when using CPU's iGPU compared to CPU only. It seems similar to using CPU's SSE or AVX units to get a speedup. Efficiency in power usage I don't know if it also gets 20% more power usage? I like the approach as currently iGPUs are idle now could be used to speedup CPU folding.

[MeeLee]

Would one CPU core be assigned to the IGP, or does the system work different from DGPU cards?

[toTOW]

The problem on these hybrid cores is that you can't know easily if it is an iGPU to use with the CPU or a discrete GPU that you shouldn't use and leave free for a full GPU core ...

Using a discrete GPU with an hybrid core would be a waste of resources : the GPU would wait a lot for work from the CPU. It's what I saw on BOINC projects that tested this solution ...

I guess that's the reason why FAH never used this kind of solutions.

[foldy]

The problem on these hybrid cores is that you can't know easily if it is an iGPU to use with the CPU or a discrete GPU that you shouldn't use and leave free for a full GPU core ...

It can be decided by GPU vendor:
* Intel iGPU is not supported by FAH OpenMM core => so use it for FAH CPU core.
* NVidia does not have an iGPU so use nvidia GPU for FAH OpenMM core only.
* AMD has both discrete GPU and iGPU. You can decide by device part number.

Similar to GPUs being too slow or not supported by FAH OpenMM core there would be a list of integrated iGPUs which are supported by FAH CPU core or not. Maybe even some discrete GPUs which are too slow or old for FAH OpenMM core could help the CPU instead.

I hope it is easy to setup by just telling a future FAH Gromacs 2019 CPU core to use an existing iGPU or not.

[MeeLee]

If I would guesstimate,
I'd say that older CPU cores with IGPs, usually have 4 to 12 cores running between 400-900Mhz.
If 1 CPU core at 2Ghz is lost to feeding the IGP, in worst case scenario you'd be trading a 2Ghz thread for a 1,6Ghz thread (4 cores at 400Mhz), and actually would lose performance compared to running both CPU cores to fold.
At best, you'd run an equivalent of about 10Ghz on the IGP (if it had 12 cores running at 900Mhz); and speeds would increase to an equivalent of 4Ghz to 12Ghz; and it would triple performance.
At first it would look worthwhile, however, older APUs would probably rise from 1-3k PPD to 3-9k PPD. I don't think it would make much sense doing the research on those.

A modern CPU, like the AMD Ryzen 5 2400G running 4 cores of 3.6GHz (or a theoretical 14.4 Ghz performance), would trade one CPU core for 11 GPU cores running at 1250Mhz, or a theoretical trade in of 3.6Ghz for 13.75Ghz; would increase performance by 70%.
Not a lot of information out there on CPU folding numbers on the new Ryzens, but many mention that the IGP's speed is a lot slower than a similarly clocked DGPU, due to the much slower RAM.

[Theodore]

The new Ryzen 3 3200G would probably be the first one I would fold on.
With 512 stream processors, clocked at 1,2Ghz, it wouldn't be strange to see scores of up to 100k PPD.
It should be supported, as the IGP runs off of AMD radeon drivers.

[Joe_H]

As long as the device ID for the iGPU in that 3200G is listed in the GPUs.txt file, it will be usable for folding. That assumes they include support for double precision. How well it will do is going to depend on how fast it can run continuously, the IGPU on that chip shares TDP with the CPU. So even with 512 stream processors, it may run at a speed much less than the 1.2 GHz possible.

[foldy]

That's the problem with iGPUs for FAH GPU folding: most are too slow for FAH which means they can block the chain. FAH GPU results are generated in a chain and FAH goal is to work through the chain fast.

That is different for FAH CPU work units which heavily run in parallel and could in theory use the slow iGPUs to speed up FAH CPU folding in a future version maybe.

[toTOW]

No, the principle is the same on CPU and on GPU : each completed WU generates the next one in the trajectory and there are multiple trajectories with variations in initial conformations.

[bruce]

No, the principle is the same on CPU and on GPU : each completed WU generates the next one in the trajectory and there are multiple trajectories with variations in initial conformations.

Saying the same thing in a slightly different way:
For almost all projects (CPU or GPU) the first 3 numbers, (Project, Run, Clone) denote a specific trajectory. The last number (Gen) denotes the segment of that trajectory. If you slow up that trajectory, either by using an extremely slow GPU or an extremely slow CPU, the next segment (Gen N+1) cannot be issued until you upload the segment your'e working on.

FAH's parallelism is created by assigning a different P,R,C to everybody else. Nobody else can work on the trajectory you're working on until the Gen you're working on has been completed or has expired.