| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: OpenPandemics - COVID-19 Project
Thread: Mandatory "Request for Project to be Ported to GPU" Thread |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 133
|
|
| Author |
|
|
William Albert
Cruncher Joined: Apr 5, 2020 Post Count: 41 Status: Offline Project Badges: 
|
Look at this list and ask yourself do all these corporate type whales have CUDA 11 GPUs on their computers? But they are doing OpenCL. As one of those corporate whales, most of my machines are servers without a GPU of any sort. |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
Well after running a few on the CPU, they average 576 seconds. And I fired up my GTX 1060, and they are doing 36 seconds there, so a ratio of 16 to 1, not as much as I recall. You can use whatever you want. The WCG/OPN may be different of course. We need to check again.
|
||
|
|
supdood
Senior Cruncher USA Joined: Aug 6, 2015 Post Count: 333 Status: Offline Project Badges: 
|
Thanks for the details, Jim.
---------------------------------------- |
||
|
|
BladeD
Ace Cruncher USA Joined: Nov 17, 2004 Post Count: 28976 Status: Offline Project Badges: 
|
Now, the cat is out of the bag! You don't want a GPU version! You want a GPU version that run only one type of GPU! Nonsense. It could be CUDA or OpenCL. I don't even know the difference. I hope it runs on a wide range of high end GPUs. My point is that office PCs in companies tend to have very weak GPUs, e.g. GTX 260 or even use the graphics in a CPU.But to have a CPU project that is working great and then waste months just to force GPU WUs to run on either a CPU or GPU is a terrible waste of time. I think if they would've split CPU & GPU we'd currently have both running. I also think CPUs make more accurate calculations than GPUs. Perhaps find hits using GPUs that get within their desired bonding range and rerun those high scoring hits using CPUs. Or they can diverge and do something very different. Well then, it has to be OpenCL, since CUDA doesn't support AMD! |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
Thanks for the details, Jim. Having looked at it in more detail, I think that when you look at efficiency (output per watt) that a good CPU such as a Ryzen 3000 series might be almost as good as a GPU. At least it is worth a check with OPN. That is my standard of comparison, not raw output. [Edit 2 times, last edit by Jim1348 at Aug 14, 2020 9:02:11 PM] |
||
|
|
robertmiles
Senior Cruncher US Joined: Apr 16, 2008 Post Count: 445 Status: Offline Project Badges:
|
Look at this list and ask yourself do all these corporate type whales have CUDA 11 GPUs on their computers? But they are doing OpenCL. In either case, the GPU would be probably at least 20 times faster than the CPU, from the ones I ran on Quarantine@Home. So the corporate whales become more or less irrelevant. They can use their CPU cores elsewhere. A lot more crunchers who are looking for GPU work would jump in to take their place. I hope it is CUDA 11. That would be best. The GPUs that can do CUDA 11 are all from Nvidia, and do not include graphical interfaces. In other words, you cannot find them on any graphics cards, only on cards that can do CUDA or OpenCL, but not graphics. The speed of a GPU version of an application varies widely, depending on the application and the number of GPU cores. It can be as slow as a quarter of the speed of the CPU version, since the GPU clock is usually about a quarter of the speed of the CPU clock. It can also be up to the number of GPU cores divided by four, times the speed of the CPU version. Some GPUs have enough GPU cores to make this maximum about 750 times the speed of the CPU version. Speeds close to these extremes are rare, partly because BOINC projects seldom release GPU versions that aren't expected to run at least 10 times as fast as the CPU version. [Edit 1 times, last edit by robertmiles at Aug 15, 2020 1:42:39 AM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Over at F@H they have had a unified work load for years. All of the WU can run on CPU or GPU. Yes, they use OpenCL to do this. It works fine and there are not bugs thanks to extensive testing.
----------------------------------------What I hope to see in the future for OPN isn't just GPU workloads, but AVX1 support with parallel processing for CPUs. That would turn any 4 core CPU since 2011 into a 100-500 GFLOP beast and your modern 8 core CPUs into TFLOP Titans! [Edit 1 times, last edit by Former Member at Aug 15, 2020 3:03:08 AM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Parallel processing on CPU is just a gimmick. It is lossy in time and reduces overall throughput... one thread waiting on the input from another thread still calculating.
As was noted, the aim is to have a single workunit that can be processed on either, or. The advantage being that if a particular result fails on GPU, it can be rerun/retried on CPU and v.v. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Parallel processing on CPU is just a gimmick. It is lossy in time and reduces overall throughput... one thread waiting on the input from another thread still calculating. I couldn't agree less. I know for a fact many kinds of workloads are sped up by parallel processing, including, but not limited too, molecular dynamics modeling (F@H, Rosetta, and others), video encoding (Handbrake), particle physics simulations (LHC@Home's Atlas@Home), ray tracing (PovRay, LuxRender, Blender), and many, many more. On the other hand it is also true that if a task performs best as a single thread then it should stay a single threaded. Tasks like these include LHC@Home's Sixtrack project, Climate Prediction.net, defragging your HDD in Windows, and many, many others as well. I say use the right tool for the job and if the GPU port is being done to leverage parallelism, which it is, why don't me leverage it on CPU as well? [Edit 1 times, last edit by Former Member at Aug 15, 2020 9:54:42 AM] |
||
|
|
KerSamson
Master Cruncher Switzerland Joined: Jan 29, 2007 Post Count: 1684 Status: Offline Project Badges: 
|
In the specific context of number crunching, multithreading is less meaningful than to compute several WUs (one / CPU thread) simultaneously.
----------------------------------------I mentioned it already several times, porting a project to GPU is not as easy as many people think it is. Data and application must be made "GPU compatible" AND the validator must be able to validate correctly GPU results, otherwise many people will complain about the high number of invalid results. From a feeder point of view, being able to maintain the same WUs for the both environments is obviously much better and flexible. I can easily understand the Tech Team strategy, trying to keep this kind of "standardisation" throughout the projects. Following a "try and error" strategy is neither meaningful nor fair for the contributors. For this reason, implementing GPU support takes time for being as highly reliable as possible from the starting point (Beta with contributor). In other words: please relax and be patient !... Cheers, Yves |
||
|
|
|