Folding Forum

We used to say that the GPU cores don't use Double Precision. That was true until recently when FahCore_21 started using a little Double Precision. The developmental target for Folding@Home has always been GPUs that are commonly found on home computers (hence ATI and NV have been targeted). FAH would also run on "professional" GPUs which generally stress Double Precision performance, though that was not used. (Home/Game GPUs are much cheaper, too, unless your boss has given you written permission to fold on the company workstation.)

Many older home-style GPUs do not support any DP. Many newer ones do provide hardware support which is not particularly fast but does work. Since the actual use of DP is a rather small portion of the analysis, speed isn't the problem, but for FahCore_21, zero support is a bad thing.

That's a long way of responding to holdingpattern's "Anandtech indicates that Intel Iris 6200 supports double-precision" which, for the most part, isn't a significant issue. If an Intel GPU is ever going to be supported by FA , a first step would be to support Core_17/_18 on a broader range of Intel GPUs, even ones without DP, before focusing on the 6200 or others which might have DP support.

I have Intel HD 5500 here, allthough I agree that we can never expect a "real" GPU performance from such embedded chips, they may have advantages with the huge amount of number of units installed.

So, I wonder 2 things. Why doesn't Intel step in and provide support&help supporting their own GPU? As an ex video professional their (unrelated) QuickSync video transcoding acceleration really impressed me. Perhaps their OpenCL 2 support may surprise people too?

Can we run some tests here, on real chip with real up and running OS? Would it help?

The traditional answers to that question have always included words like "integrated GPUs are really slow."

To associate some additional information with those words:
Based on a recent test om a Mac Laptop in Stanford's lab, it turns out that folding identical proteins two different ways, with the iGPU and with the CPU, the CPU was faster. In other words, creating the infrastructure to assign projects to an integrated GPU (which, in turn, relies on the CPU hardware for processing resources) isn't worth doing. The project already runs more efficiently on the CPU directly.

The real benefit of processing on (discrete) GPU hardware is that the computational circuitry provided in the external GPU chip is much more powerful than on CPUs.

I have not "done my homework" with recent backup data, but I suspect if you can find GFLOPS numbers that allow a direct comparison of your Intel CPU and its iGPU, you'll probably find it supports this test result. Then, too, the efficiency of compiler-generated code for a CPU is probably somewhat better than for the GPU, based on the respective number of years of compiler development.

It would seem that porting FAHBench would be the easier way to go, just to "get some numbers."

Intel Graphics support Single and Double Precision math.

A typical Skylake-powered laptop, is going to have a decent amount of computing power inside a very efficient power envelope that can do between 100-500 Gflops, that's a lot of double precision math to be ignoring on consumer level equipment.

The Intel 520 has 24 EUs, calculate DP performance as follows @ 2.3 GHz

EUs * 4 [SIMD-4 FPU] * 2 [MUL + ADD] * clock speed / 2 = 221 DP Gflops on a garden variety i5 laptop.

Welcome to the folding support forum.

You reopened this topic with an example based on the next newer generation of integrated video available for Intel chips. Many of the same factors apply before they could be used for folding, drivers with sufficient OpenCL support from Intel and other items mentioned in previous posts in this thread.

However, though you can calculate a DP performance value, what is the actual Gflops available? Typically for consumer level graphics that is a small fraction of the computed value. In addition, with only 24 EUs that places a large limitation on how much GPGPU processing that can be done. The effective GPU's typically have hundreds of usable processing units available. The low end cards are not very effective, those with few shaders often can not complete folding work within applicable deadlines.

Good afternoon Joe,

Perhaps a good opportunity to start another thread?

With our own benchmarks, and those available publicly, with OpenCL and Direct Compute, show particle simulations with 64k atoms, for example, show performance equivalent to a Geforce GTX 590, which is rated by the NVidia at 365 DP Flops.
So, there is evidence, contrary to your opinion.

They're not going to replace a GTX1070, but they can compute about 1/6th to 1/7th of that amount, so an Intel EU can do about 8-16x the flops of a Cuda core.

Also, Intel's SDK for OpenCL, seems to be updated regularly. The most advanced Intel iGPUs, the Iris 580have 72 EUs, and EDRAM, putting them 1/4 the performance of a 1070.

I think the time is here. We're considering working with Intel, on clusters of E3s with Iris graphics, for Gromacs and Amber.

I think that folding on the igpus is to earlie reason being when you are cpu folding your chip will get hot but when you are also going to use the igpu chip that is placed on the cpu chip things will get even hotter and this can cause thermal throttling and then igpu folding isnt worth it at all and you are going to find your limit really fast if you overclocked your cpu so for now I think it is to earlie if mores law continues then it might be possible one day but that are just my thoughts on this toppic

I read that they tried running OpenMM, the folding core, on intel iGPU but it wasn't worth it with not iris CPUs. So they went for CPU AVX. But maybe in the future.

When we see 65000 active Windows CPUs and guess that 10% would run iGPU efficiently now or in the future then this are 6500 iGPUs.
If we assume 0.25 Tflops for these then this is a potential of 1600 Tflops. Similar to the potential of a Mac OSX gpu client.

I tried FahBench 2.2.5 on my intel i5-2500k but it fails with "Potential energy error".
Edit: Using Windows 7 64bit
I dont't know if it even was run on iGPU or the CPU part which also can do OpenCL.
Looks like on my Sandy Bridge CPU the Intel iGPU driver does not support OpenCL.

If FAH managment ever decides to support Intel iCPUs, it's going to be a result of actual reports like foldy's. If Donor reports demonstrate that potential of 1600 Tflops is a realistic number then FAH can make a more realistic assessment if the extra development work is worth the effort.

1) Are the errors (like his "PE Error" a result of Intel bad drivers on (Windows/Linux/MacOS} is there an upgraded driver version available?
2) Is there something in FAHBench V x.x that needs work?
3) If we can get successful FAHBench performance reports we can then categorized them by OS and by CPU model.
4) Does running FAHBench on the iGPU reduce the performance of the CPU client (with or without AVX support)?

The only way to demonstrate it's worth funding a development project is to gather enough facts to counter the negative impressions based on previous testing.


I'm searching for information on the status of Intel support in FAHBench.
If reports

bruce wrote:If FAH managment ever decides to support Intel iCPUs, it's going to be a result of actual reports like foldy's. If Donor reports demonstrate that potential of 1600 Tflops is a realistic number then FAH can make a more realistic assessment if the extra development work is worth the effort.
...

Imo, it would only seem logical that whatever the "actual" is, after 9+ total generations (over nearly 7 years between PC and notebook) it's obvious that the 21st Century includes this tech and as such is "worth the effort" to include the millions of units sold.

With every question, about the software, comes the reply of "cost....cost....cost..."; at some point, the counter has to be accepted: "If you want the project to grow, the software has to include more hardware".

Just my 2¢

Cost is simply one key measure for comparison. Hiring good programmers isn't cheap.

FAH recently decided that improving the CPU client by adding AVX support was a better investment than other thnings that were considered. It's possible that they ran into issues like Foldy's "Potential energy error" (and fixed them) or maybe not but they got that support as far as being used on a limited number of projects. It's possible that there's more benefit from AVX support than from iGPU support or it's possible that isn't true. All such considerations were probably included in the decision and FAH went forward with a development project. This also had the advantage of bringing a new version of GROMACS to the table. (I don't understand the details of what that means to science other than the obvious production improvements in the CPU core.)

bruce wrote:Cost is simply one key measure for comparison. Hiring good programmers isn't cheap...

I don't deny that (being one, I'm well aware of the charges); my point is that, sometimes, nominal improvement x becomes 1000x to 1000000x when it opens the science to 1000000 more potential donors.
Even if HardwareY is only capable of producing 1 WU, 1000000 units of HardwareY would produce 1000000 WUs that will otherwise go undone (which is about a 1:11 ratio of all WUs ever produced by the entire Overclock.net team).

Again, just my 2¢

The iGPU on an intel chip is already supported by the same drivers that support AMD's dGPUs, so the potential donor count would have to exclude AMD's CPU ... i.e.- ONLY include true Intel machines.

I also expect that Intel's iGPUs aren't all that different than AMD's iGPUs at least by some large factor. Most people with AMD systens quit folding on their iGPU because their system is too slow to meet the deadlines. Actual numbers from FAHBench would be more meaningful than this estimation but somebody needs to provide those measurements to support your enhancement request.

I think this all boils down to a question of whether the Broadwell drivers provide double precision support through OpenCL.

The Energy calculation that failed for Foldy suggests that his drivers were not supporting Double Precision. If the capabilities of Broadwell are not exposed through OpenCL, FAH can't use them.

@Foldy: Do you have the OpenCL 2.0 drivers installed?


To everybody else:
Does anybody have a machine with the OpenCL 2.0 drivers install that can run FAHBench tests?

What percentage of Intel machines will have it installed?