Folding Forum

Now that the new Intel Ivy Bridge processors are out and the HD 4000 supports OpenCL 1.2 (I believe) is there any plan to support it for calculations? I have no idea how much it would add, but my philosophy is every little bit helps! (Which is why it's still worth running the CPU despite all the computing power the GPUs deliver).

try adding it. GPUs.txt

Please stop recommending that people add stuff to the GPUs.txt file. That doesn't do anything. It has to be changed at the PG servers first, then added to the Stanford copy of the GPUs.txt file, then you can download a fresh copy, whenever that happens. Then you can install and run it.

And that Intel chip is unlikely to be supported, at least not for a while.

Understood. Here's the GPU-Z dump of it anyway:



Thanks

GPU Z isn't reporting direct compute or open cl support on even some of the new ATI/AMD 7xxx series gpus..... so it's by no means a reliable way to determine proper opencl support.

Intel's HD4000 however does do fairly decent openCL number crunching.. mind it it's not amazing.. it however does do as good a job as the currently lowest supported HD5450 for example. Which is about a 2 day fold time. It would be nice to get support for the intel HD4000, i suspect the 3000 and 2500 series would be able to fold within the alotted time completeing it before the deadline. Although the HD2000 would probably ride the usual deadline. Not sure... obviously being it's own seperate gpu..... the point/deadline/wu may be talored for it. However since everything is going OPENCL now.... I think F@H should start implementing a uniformed WU system for OPENCL ... set deadlines and then present bonuses on completition times similare to that of CPUs. If the HD4000 proves to be to slow to complete WU before the deadline, then so be it.

During early testing, some folding was attempted on HD4xxx hardware. At that time, Development determined that those GPUs were too inefficient when given the type of workload that FAH required to be worth it. It would be more efficient to use the CPU time by adding a Uniprocessor WU or by releasing that CPU time to be used by SMP. Some important GPU hardware was added at the HD5xxx level.

bruce wrote:During early testing, some folding was attempted on HD4xxx hardware. At that time, Development determined that those GPUs were too inefficient when given the type of workload that FAH required to be worth it. It would be more efficient to use the CPU time by adding a Uniprocessor WU or by releasing that CPU time to be used by SMP. Some important GPU hardware was added at the HD5xxx level.

That HD4000 is not that HD4000 - Intel and AMD GPU naming is somehow the same. As the OpenCL is not the same everywhere there is likely need for a new GPU core for the Intel. Is anyone at the PG willing to make one?

Oops. I saw HD4000 and missed that he specified Intel. Forget what I said.

The first GPU core used DirectX. It didn't do very well due in part to lack of support from the vendors. Later ATI and Stanford teamed up and developed GPU2 which was a significant improvement. More recently, CUDA and OpenCL have come along and progress continues.

One key question: How much support will Stanford or OpenMM get from Intel?

Others: Does Stanford have a development budget that can support the necessary improvements to the existing FahCores plus a development project for a new FahCore? How much of a new core will be identical to other OpenCL cores and in what way might it need to be unique? Will the total amount of folding processing would be accomplished on Intel and how does that compare to investment required to develop a new core?

Intel does have very large share of the CPU market and very large % of those CPUs does have integrated GPU. Also, Intel is actively developing their drivers and OpenCL support including Linux side of things. Maybe they are willing to contribute some time and effort for the OpenMM development as well. Someone from the PG should ask for it.

Can´t imagine how much development is necessary for the Intel Hardware. But it meets the OpenCL-standards and shows a phantastic performance, look at CLBenchmark results. eGPUs are fast, but this iGPU is efficient, IMO fast enough. Yes i understand as far as PG gets enough GPU power from eGPUs there is no need to develop.

But what a gain for both, Intel and PG, if they would cooperate.

P.S. Waiting für BOINC-support, it will be supported the next weeks/months i hope, see Changeset 25792 in boinc.

Latest 15.31 series beta driver supports OpenCL 1.2 now also :-



It's quite a capable IGP to be honest, I'm sure it could do some decent work.

Any news on the HD 4000 support, is this something that's in the pipeline at least?

I suppose it's fair to presume that it's somewhere in the pipeline but there's no new information except whatever you can deduce from reading viewtopic.php?f=16&t=23434.

Proteneer is developing a general purpose benchmark which can somehow harness a variety of CUDA or OpenCL software to a variety of hardware (including CPUs as well as GPUs) and process enough of a couple of proteins that it can report comparative benchmark numbers. The steps being worked on in that topic are a prerequisite to developing reliable FAHCores. There are still quite a few additional steps that are critical to establishing that a particular combination of GPU/Driver/OpenCL/FahCore is ready for validation as a dependable scientific tool.

A FahCore and a whitelist entry is still at some vague point in the future.

Thanks for the info Bruce, I guess that's fair enough.

On just about any CPU that has the intelHD 4000 you're going to get better folding and less lifespan reduction of the cpu simply by doing SMP loads. The 4000 is roughly equivalent to an ATI 3780 in total power and relies heavily on the CPU for much of its function. the OpenCL is just there for compatability to be honest.
(Even the 8670D core on the new FM2 APU have very very little umph: More power in a half-clocked 6950.)
I wouldn't trust the poor little intel component to last long.