To better reflect the nature of the application, the AMD GPU application is now called Molecular Dynamics for CPU and AMD GPUs.

It will be issued to any machine that:

* accepts CPU work
* has no NVIDIA GPUs
* has zero or more AMD GPUs

The intention is that it will use CPU cores, and be further accelerated by an AMD GPU is one happens to be present.

Have you found a way of warning the BOINC client that the AMD GPU will be used, so that it can schedule other projects around it? Might be a bit messy otherwise, especially if the other project expects to use a high proportion of the GPU's resources.

BOINC 7.3.11 immediately designates the 9.17MD task to high priority for CPU only. 100% CPU usage with task reporting progress showing 2 1/2 hr. estimate until completion. (10% computed in 16 minutes)
Thank you for opening the MD app to CPU only machines.

For Linux, will these jobs still run on a multi-core AMD CPU SMP system without an AMD GPU?

I have a Ubuntu 14.04 system with 4 hex core AMD CPUs, on which I've run these jobs well in the past. I've now added 6 Nvidia GPUs that are running the long ans short GPU jobs. The system will run the Milkyway@home's OpenMP tasks using 16 CPU cores while GpuGrid jobs run on the NVIDIA GPUs, but I cannot get it to run the GPUgrid multi-CPU jobs now.

When the system fetches new work, it is only get GPU tasks. Can you run GPUgrid jobs on both CPU and GPU on the same system?



Thanks,
Greg

For Linux, will these jobs still run on a multi-core AMD CPU SMP system without an AMD GPU?

Greg, nothing for Linux yet. For the moment there are too few Linux machines with AMD GPUs attached to justify the effort.

Matt

I asked it you still support CPU jobs on systems without a GPU on Linux

I used to be able to run your jobs for Molecular Dynamics on my CPUs on Linux. Each job would run really fast using all 20 cores. I've got a small cluster with 168 AMD Opteron CPU cores, that these apps used to work on, but I noticed that I have not gotten any work on these for 2 months. The cluster runs over a million cobblestones per day, but less and less of it seems to work on GPUgrid now.

After you dropped support for my older Telsa cards, I've got them working on MilkyWay jobs. Only the head of my cluster has a GPU that is new enough for GPUgrid jobs, but the 6 headless nodes could ran CPU tasks. Now my CPUs can't get work from GPUgrid either.

Will the Molecular Dynamics for CPUs work again on Linux, or should I drop the GPUgrid projects altogether on those machines without a GPU?

Thanks,
Greg

Greg,

There won't be any more MD tasks for Linux CPU clients for the foreseeable, but there will be a new multithreaded application along presently specifically for those.
Please don't detach your clients!

Matt

A question, please, regarding

Molecular Dynamics for CPU and AMD GPUs.

The intention is that it will use CPU cores, and be further accelerated by an AMD GPU is one happens to be present.

this desktop:

2015-03-20 | CUDA: NVIDIA GPU 0: GeForce GT 640 (driver version 347.52, CUDA version 7.0, compute capability 3.5, 1024MB, 606MB available, 803 GFLOPS peak)
2015-03-20 | OpenCL: NVIDIA GPU 0: GeForce GT 640 (driver version 347.52, device version OpenCL 1.1 CUDA, 1024MB, 606MB available, 803 GFLOPS peak)
2015-03-20 | Processor: 8 GenuineIntel         Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz [Family 6 Model 42 Stepping 7]
2015-03-20 | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 sse4_1 sse4_2 popcnt aes syscall nx lm avx vmx smx tm2 pbe

can run 8 processes

Why are MD WUs prohibited...

It will be issued to any machine that:
* accepts CPU work
* has no NVIDIA GPUs
* has zero or more AMD GPUs

...from being sent to PCs with NVIDIAs ?
Is this some programming limitation or bug?
Thanks.

I agree, that this application should be sent to a host, even if that host has NVIDIA GPUs and 0 AMD GPUs.

Why can't we do this?

Why are MD WUs prohibited...

Because your machine has an Nvida GT640 GPU.

Matt

I agree, that this application should be sent to a host, even if that host has NVIDIA GPUs and 0 AMD GPUs.

Why can't we do this?

Because I want the machines that can be running ACEMD to be doing so. That work is worth more to us.

Matt

But I've got 8 CPUs that can also be helping you!!!

I think you're thinking about this wrong. We're looking to give MORE resources to you, NOT LESS.

But I've got 8 CPUs that can also be helping you!!!

I think you're thinking about this wrong. We're looking to give MORE resources to you, NOT LESS.

There are additional limitations of the application and scheduler that mean it's not currently possible to drop the app onto a machine with an Nvidia GPU and tell it not to use it.

Some other project will have to benefit from our short-comings.

Matt

Just to clarify... is the behavior "I will use an NVIDIA GPU and you can't tell me not to" ... a behavior/limitation of the application executable, or is it a behavior/limitation of BOINC scheduler/API (which David A might be able to fix)?

Just to clarify... is the behavior "I will use an NVIDIA GPU and you can't tell me not to" ... a behavior/limitation of the application executable, or is it a behavior/limitation of BOINC scheduler/API (which David A might be able to fix)?

It's both the application, which will blindy use any and all GPUs it finds on a host, and also the old version of the BOINC scheduler we have in production on GPUGRID, which knows not of OpenCL applications.

Matt

Thank you, Matt. Those are the answers we were looking for, I believe. Have a great weekend, and if you feel like tackling a bug, there's a power issue in the Number Crunching forum :)

So, it seems my PC is not allowed to run 'short runs' or 'long runs' applications on its NVidia GPU while also running "Molecular Dynamics for CPU" application WUs on its 8 CPUs.

Per Jacob, we are trying to offer more computing power to you, not less.

Is this a constraint or limitation on GPUGRID (as a BOINC project) which may be removed in the near future, or no?

Thanks.

---

I think ∴ I THINK I am
My thinking neither is the source of my being
NOR proves it to you
God Is Love, Jesus proves it! ∴ we are

So, it seems my PC is not allowed to run 'short runs' or 'long runs' applications on its NVidia GPU while also running "Molecular Dynamics for CPU" application WUs on its 8 CPUs.

The opposite - if your machine can run "short" or "long", that's all it will get. These are our main applications and contribute most to our scientific work.
The other applications are of secondary importance, and I don't want them cutting into the return from short & long.

Matt

And what about Intel GPUs?

Even they are not the most powerful latest Iris GPUs are comparable powerful to 2-3 years old top Nvidia GPUs, 2-3 times more powerful then CPU, more power efficient then CPU, are available on all desktop and mobile Intel CPUs and their computational power is on top of CPU one...

I think with adoption of OpenCL for AMD GPUs it may be relatively easy to include support of OpenCL for Intel GPUs as well.

Note: It would be better to add AMD to supported platforms on Boinc project page where people are make first decision what project to support.

So, will there ever be an WU generated for CPU or AMD GPU???
Otherwise goodbye