Intel® HD Graphics

[ProDigit]

I currently am in the process of installing an Intel Xeon system, 12 cores 24 threads, with a single Nvidia GTX1050.
Price per performance was the best on this card.
the GTX 960 was going for $55-60,
The GTX1050 was going for almost the same price
The 1060 was going for $260,
And the 1080 was going for over $500.

The 960 and 1050 are almost identical in performance, but the 1050 uses 50% lower TDP (75W vs 140W); which in the long run, will pay back on electric bills.

[Theodore]

Perhaps time to revisit this topic.
Intel IGPs support OpenCL, and the new Icy Lake CPUs (15Watt TDP) have been reported to have 1.12 teraflops of FP32 compute power. Their higher TDP versions are more than likely going to perform even better.

While this isn't very much (about the same as a GT 1030, averaging around 35k PPD), their performance comes from the sheer numbers of people owning Intel graphics on their PCs, many of them currently folding on CPU only.

It could also mean that many Intel CPU based PCs running AMD or NVidia cards for folding, will see an additional boost in folding, provided multi brand GPUs can run on one system without issues.

[darkbasic]

They also have a brand new OpenCL stack which is very good.
But it won't be enough for them to whitelist it.
Eventually they will, but it would only be thanks to the discrete Intel GPUs.

[Joe_H]

What is their FP64 performance? If they do not support double precision, they are of no use to current GPU processing as all projects going forward are using mixed precision calculations.

As for OpenCL support, this is nothing new for the Intel iGPU's. One issue that has existed in the past when they tested it was that it's implementation of OpenCL was not sufficient to support folding. Unless that has been much improved, still not going to be usable.

[darkbasic]

Have a look at the NEO OpenCL driver.

[MeeLee]

It was my understanding that FP64 (double precision) isn't being used (much) by F@H. Why would they change their system to FP64?
If that's the case, a lot of people using RTX cards will be disappointed they did, as AMD does FP64 a lot better.
And scoring systems(PPD) will change too.

Not much is known about Icy Lake's double precision floating point performance, other than we can expect them to do less than 5/10ths the FP32, to the worst being only 2/10ths the FP32 performance.
At least we know their performance is 1/4th to 1/3rd better than previous generation.

[JimboPalmer]

It was my understanding that FP64 (double precision) isn't being used (much) by F@H. Why would they change their system to FP64?

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

If the intel GPUs support OpenCL 1.2, then all that is needed is someone with deep pockets to fund development of a Core.

[bruce]

Not much is known about Icy Lake's double precision floating point performance, other than we can expect them to do less than 5/10ths the FP32, to the worst being only 2/10ths the FP32 performance.
At least we know their performance is 1/4th to 1/3rd better than previous generation.

The comparison chart you posted is not applicable. (It does describe Gaming performance.) Good gaming performance demands fast, high-resolution graphics ... but FAH does not use ANY of the graphics hardware. (See bold items below.)

The slowest Graphics device is more than enough to display your desktop which is all that's needed if your are using FAH on a computer that's also doing other things, excluding games. AMD and NV discrete GPUs do provide good graphics performance since they're aimed at gamers but that portion of the GPU is not used by FAH and spending money to get improved frame rates is a waste if you're goal is FAH.

Looking at the Sensors tab of GPU-Z (with zero effort on my part to optimize for FAH)
GTX 750 Ti
Fan Speed = 79%
Memory Used = 182 MB
GPU Load = 99%
Memory Controller Load = Variable (19 - 30%)
Video Engine Load = 0%
Bus Interface Load = 37 - 40
PerfCap Reason: Thm

GTX 960
Fan Speed = 34%
GPU Load = variable (97 - 99%)
Memory Controller Load = 59%
Video Engine Load = 0%
Bus Interface Load = 22%
Memory Usage = Dedicated 65 MB / Dynamic 43 MB
Power Consumption - variable (87 - 112 W)
PerfCap reason - None

[bruce]

It was my understanding that FP64 (double precision) isn't being used (much) by F@H. Why would they change their system to FP64?
If that's the case, a lot of people using RTX cards will be disappointed they did, as AMD does FP64 a lot better.

Let's say that in a typical WU, 90% of the calculations can be done with FP32 and maybe 5% of the calculations require FP64 precision. (The actual numbers will depend on how many atoms are in the protein and a few other factors). Not having any FP64 capability means you cannot process those WUs. Having fast or slow FP64 doesn't matter much since it's the speed of FP32 that matters.

In the early days of FAH, proteins being studied were significantly smaller and F32 was sufficient. Gradually the science that needs to be done today has replaced those studies that could be done with FP32, just as 64-bit versions of an OS have replaced 32-bit (or even 16-bit) OS although there still are a number of programs which can be run in 32-bit mode.

[foldy]

The latest gromacs supports GPU too as helper for CPU. So a FAH CPU core could use the intel GPU as accelerator when using the latest gromacs.

[JimboPalmer]

The latest gromacs supports GPU too as helper for CPU. So a FAH CPU core could use the intel GPU as accelerator when using the latest gromacs.

{The following is wild speculation, and should not be confused with Reality}

It would make a lot of sense if Intel's software developers wrote the Core for Intel hardware. They know the quirks, and strengths and weaknesses of the driver.

Intel management will not wish to do a lot of work to come in third in a 3 way race. Your idea of augmenting CPU work could hide the fact that they came in third, allowing them to present it as a boost to CPU WUs, rather than weak GPU WUs.

In reality, I do not expect Intel help until they have competitive, discrete hardware. I cannot think of another deep pockets contributor who would want to make Intel look bad. (Aside from Nvidia and AMD)

[foldy]

Intel is not needed to help for that. Latest Gromacs supports CPU added iGPU usage out of the box. And some Intel CPUs also bundles AMD iGPUs now, so there the GPU core already can run.

[bruce]

I like that, Jimbo. Do you mind if I quote you?
{The following is wild speculation, and should not be confused with Reality}

I have not researched the latest version of GROMACS, but is it encouraging. I do have a question.

FAH is likely to upgrade their CPU core (Core_a7 -> Core_a(8 ?) when something they consider essential is incorporated by Gromacs. How does gromacs determine if a OpenCL device is a iGPU or a dGPU?
If a new version of Core_a* steals resources from a dGPU, slowing down the GPU slot(s), they're like going to disable that feature.

Traditionally, OpenCL from Intel was full of bugs and limitations and I'm afraid those facts have prejudiced FAH's development team -- having spent time and money trying to get a an Intel core to work. (Also, the iGPU is orders of magnitude weaker than the dGPUs that are supported.) They've got other priorities to spend their development money on that are more fruitful. Even just trying to validate the new OpenCL is pretty far down their list. This has put Donors who run on the Mac platform at a distinct disadvantage since most Macs don't have AMD/NV dGPUs. Those guys with Macs would like to see the same things you're asking for.

Maybe you can convince Intel and/or Apple to donate some money to FAH Development -- or to fund the entire development effort.

[JimboPalmer]

Intel is not needed to help for that. Latest Gromacs supports CPU added iGPU usage out of the box. And some Intel CPUs also bundles AMD iGPUs now, so there the GPU core already can run.

if it worked like you think, we would have it now.

I bet there is going to need to be development time/expense/expertise.

[bruce]

My question still stands. How is the FAHClient going to tell if the AMD GPU is an iGPU or not. They won't want the WU assigned to a CPU slot to be stealing resources from a GPU slot -- but stealing resources from an underutilized iGPU would be a good thing. that complicates the logic FAHClient has to use before upgrading from the gromacs version that incorporated AVX to the version that incorporated the parallelism features of an iGPU.

Do we know which provides better FP parallelism if both are available: AVX or the IGPU support?
(On a non-AVX CPU, the question becomes which provides better parallelism, SSE2 or iGPU support?)

i.e.- across the entire spectrum of AMD or Intel CPUs, which provides the greatest degree of FP parallelism.