some sticky info for gpu
 

I have been following mersenne forum forever, now with this new GPU section, it is a bit difficult to find information to get started...

Also on gimps first page, download software of getting started says nothing about gpu...

Would it be possible to make sticky discussion, how to start, what are the pre-requisities, linux, wintoys....

At least today, I dont find it enough easy (quick) to find instuctions for linuxbox setup...

Will come
 

It is [URL="http://www.mersenneforum.org/showpost.php?p=244237&postcount=494"]planned[/URL]. Job simply needs to get done (by someone).

Maybe we can use this thread to do the planning and organizing for the stickies.

One will be needed for "Current versions of the various software".
Another for the "How to TF on a GPU".
Another for the "How to LL on a GPU"
Maybe another for supported and unsupported cards.

Also, someone like [URL="http://www.youtube.com/watch?v=BkdOm5BJesk"]Luigi[/URL] (ET_) or Oliver (TheJudger) should be made a mod of this sub-forum.

[QUOTE=Uncwilly;244367]Maybe we can use this thread to do the planning and organizing for the stickies.

One will be needed for "Current versions of the various software".
Another for the "How to TF on a GPU".
Another for the "How to LL on a GPU"
Maybe another for supported and unsupported cards.

Also, someone like [URL="http://www.youtube.com/watch?v=BkdOm5BJesk"]Luigi[/URL] (ET_) or Oliver (TheJudger) should be made a mod of this sub-forum.[/QUOTE]

I'm willing to help, but Oliver has a better knowledge of the whole system :smile:

Luigi

Better yet, just add GPU support to Prime95 already!

well... lets say i have one linux machine downstairs. cpu is doing mprime. Machine hosts this graphic card: nVidia Corporation NV44 [GeForce 6200 LE]

Any chance to start with this? Might be a bit old....

In total I have now 8 computers in gimps, I would like to check all graphics cards on those machines and see can they help.

Ati I already undestood to forget, like the one in this machine :S

All machines I use are linux boxes, maybe they share similar instructions?

Hi,

[QUOTE=timsu;244612]well... lets say i have one linux machine downstairs. cpu is doing mprime. Machine hosts this graphic card: nVidia Corporation NV44 [GeForce 6200 LE]

Any chance to start with this? Might be a bit old....[/QUOTE]


- CUDA is supported on all Geforce 8 and newer cards.
- my TF code "mfaktc" needs compute capability 1.1 or better
- msft's LL code "CUDALucas" needs compute capability 1.3 or better

So for my TF code you can use all Geforce 8 series or newer GPUs except these compute capability 1.0 cards: 8800 GTS 320, 8800 GTS 640, 8800 GTX, 8800 Ultra (and their Tesla and Quadro variants)
For msft's LL code you'll need a GTX 260, GTX 275, GTX 280, GTX 285, GTX 295 or a 4xx series or a 5xx series GPU (and their Tesla and Quadro variants)

A 8400 GS would run my code but performancewise it doesn't make sense. I'm not sure how well e.g. a GT 420 would run msft's code.

Oliver

Yes, please -- a stickied how-to overview would be much appreciated. I've been here for years but have my GPU running F@H because it was simple to figure that out, but I haven't yet figured out how to run TF. I'm sure I'm not the only one. With some gentle guidance many of us can get into using our GPUs, but of course it would absolutely need to be an integral part of Prime95 to get any kind of mass adoption.

After bragging about how I can't figure out how to do GPU-TF, I went searching for mfaktc, and it's running, but making my system somewhat unusable... is this normal? It's only using 100% of one CPU core, but the whole system has huge lag now, which I've never seen with any distributed computing project before. 8800GT, i7-920.

Hi James,

the system is laggy not because one core is busy... it is busy because the GPU-cores are fully loaded (long running kernels, asynchronous memory copies, ...).

You could try to "oversieve" a little bit by adjusting SievePriemes and SievePrimesAdjust in mfaktc.ini. If you force a little bit "too much" sieving (CPU can't keep the pace with GPU) the GPU will be idle after each kernel. If you are able to compile from source you can additional reduce the kernel runtime by lowering THREADS_PER_GRID_MAX to e.g. (1<<18).
An other option to try is setting NumStreams (mfaktc.ini) to 1. The performance impact should be bigger than the other two things to try.

Oliver

P.S. source-tarball contains a little README which might be included in the precompiled binaries.

[QUOTE=James Heinrich;244919]After bragging about how I can't figure out how to do GPU-TF, I went searching for mfaktc, and it's running, but making my system somewhat unusable... is this normal? It's only using 100% of one CPU core, but the whole system has huge lag now, which I've never seen with any distributed computing project before. 8800GT, i7-920.[/QUOTE]
To elaborate on what Oliver said, the reason for the sluggishness is that operating systems don't have robust schedulers for GPUs like they do for CPUs.

On the CPU, your operating system will automatically rotate between multiple processes wanting CPU time in fraction-of-a-second slices. If, however, one process is of a higher priority than the other, the higher priority will always get precedence in the scheduler. This is why Prime95 is always set to lowest priority--it only gets whatever CPU cycles are not being used by other programs, and thus doesn't get in the way of ordinary system operation.

On a GPU, by contrast, there is little or no scheduler in place to make applications "play nice" with GPU time. Normally this isn't an issue if the GPU is only being used for graphics; the operating system's drivers figure out exactly what should be on the screen at any given time, and just pass that to the GPU for rendering. But when a GPU is used for non-graphics computations, it's a free-for-all as to which process gets the GPU time it wants (in this case, the OS's video drivers vs. mfaktc). IIRC, there may be some kind of rudimentary scheduler in most operating systems for handling GPU time, but it's not nearly as robust as the CPU's, and thus sometimes when you're running mfaktc the OS has a hard time displaying screen content normally. (It's the reverse of most incidences of system sluggishness: the computer itself is running at normal speed, it just can't display it to you fast enough.)

mfaktc is designed to run part of its algorithm on the CPU and part on the GPU. When the CPU is fast enough relative to the GPU, the GPU's part of the algorithm becomes the bottleneck, so it is maxed out and there's no cycles left for graphics display. By adjusting the sieve constants in mfaktc.ini as Oliver suggested, you can shift the algorithmic balance more from the GPU to the CPU, thus leaving the GPU slightly underutilized and more free for regular computing tasks.