Page 4 of 4

Re: Intel® HD Graphics

PostPosted: Wed May 29, 2019 8:37 pm
by foldy
First question: I would assign AMD iGPU always as GPU slot and only Intel iGPU for CPU acceleration.

Second question AVX vs iGPU: FAHBench

Re: Intel® HD Graphics

PostPosted: Wed May 29, 2019 9:32 pm
by bruce
What are the ID strings for the AMD GPUs and their GFLOPS.
Has anybody tested them on FAHBench?

Re: Intel® HD Graphics

PostPosted: Wed May 29, 2019 11:57 pm
by Theodore
AMD IGPs are significantly slower than a GT 1030. Newer Intel IGPs are faster than older AMD IGPs.
GPU folding on an Intel IGP more than doubles the PPD score, of just CPU folding, provided it has sufficient cooling (which most aftermarket coolers provide).

Re: Intel® HD Graphics

PostPosted: Thu May 30, 2019 4:33 am
by bruce
Both the (i)GPU and the CPU are capable of doing 3D FP32 calculations. SSE2 calculations are about double those on the original FPU hardware and AVX increases it again. A very fundamental question is how the 3D calculations on AVX compare to the same 3D calculations on the iGPU.

If I were the Intel chip designer, I would probably use the same hardware real-estate to do AVX and iGPU calculations so the might actually produce the same GFLOPS when only one of them is busy doing FP32. Can somebody please test that for us?

Re: Intel® HD Graphics

PostPosted: Thu May 30, 2019 7:47 am
by foldy
Intel CPU has its own AVX units which are independent of an iGPU bundled or not.
Isn't there a gromacs benchmark for that? Gromacs avx vs OpenCL intel iGPU?

Re: Intel® HD Graphics

PostPosted: Fri May 31, 2019 7:47 am
by MeeLee
@Bruce, not sure if I understand your question, however,

For 2019 we're expecting VEGA 8 and up GPUs built into mobile processors, that will surpass 1Tflops FP32, according to Wikipedia.
Mobile, VEGA 8 and 10: ... o%22_(2019)
and Desktopm VEGA 8 and 11: ... e%22_(2018)

Vega 8 is stated to have 512 cores, at 1100Mhz,
Vega 10, 640 cores @ 1,4Ghz.
Vega 11, 704 @ 1,25Ghz.

These graphics cards are faster than a GT1030, and could aid CPU folding by doing GPU WUs.
While Intel only has their future (non-released) IGPs at 1+ TFlops.
However, if a driver can recognize Intel IGP graphics, and is backwards compatible with older Intel IGPs, I'm sure that a lot of current Intel CPUs with integrated graphics, would benefit from the graphics core boost.

Additional sources: ... 178.0.html

JimboPalmer wrote:The operative word is much. GPUs which cannot do Double Precision (FP64) are not supported, even though use of Double Precision is minimized.

I guess the question is if the new intel IGPs can do FP64, and it's a good one!
Their FP32 performance is pretty good!
But as far as Intel graphics cards go, they usually only have the things they need. And nothing more.
There's a chance they won't get optimized for FP64.
But what prevents the CPU from doing those computations, and leave the FP32 computations for the IGP?

Re: Intel® HD Graphics

PostPosted: Fri May 31, 2019 3:41 pm
by bruce
As I understand it, OpenCL 1.1 support is insufficient while 1.2 or greater does support FP64, so it looks like Vega 8/11 should be fine. We can cross that bridge when we get to it.

The old versions of Intel's iGPU should be excluded, of course, but I think they already are. Devices identified as Vega will be (or already are) included in GPUs.txt and recognized by FAHClient.

The only potential unsolved problem is the hybrid GPU which switches between a supported GPU and an unsupported iGPU to save power when the driver thinks the graphics activity is light enough that it doesn't need to be in gamining/folding mode.