Is ATI HD 6570 capable to run 2 WUs simultaneously?
2 May 2013 18:25:47 UTC
Topic 196936
(moderation:
I changed the "GPU utilization factor of BRP apps" to 0.5 in the Einstein project preferences and downloaded new WUs, but my HD 6570 (2048 MB RAM) run 1 WU only. Worked WUs is visible as (0.5 CPUs + 0.5 ATI GPUs), so I think it's all right, but what is wrong? Are the ATI HD 6570 capable to run 2 WU simultaneously?
P.S. Sorry for my uncertain English.
Is ATI HD 6570 capable to run 2 WUs simultaneously?
)
Oh, sorry, I just reloaded BOINC manager and now all OK, 6570 GPU run 2 WUs. But it's interesting - how to know which GPUs models (for various AMD and NVIDIA) is capable to process which numbers of the WUs simultaneously and how to determine which number of parallel WUs is enough and optimal for GPU?
There is no (easy) way to set
)
There is no (easy) way to set different utilization factors for different GPUs... So, all the GPUs in the system will use the same setting. In your case you should see both GPUs running 2 WUs each... (of course, this is valid only if you have your Einstein prefferences set to allow work on both, AMD and Nvidia)
RE: There is no (easy) way
)
Thanks, but it's not answer for my question - where can I see the manual about a possible number of simultaneous WUs for various GPUs.
RE: RE: There is no
)
There is no real manual for this. The closest thing would be the pdf maintained by dskagcommunity in the thread "CUDA and openCL Benchmarks" over in the Cruncher's Corner of the message boards. To get the best result for your own machine and card you'll have to test it yourself.
To test it your self first get a program that can report how the GPU is used eg. GPU-Z for Windows and then try different number of task simultaneous, start with 2 and then go to 3 and so on, when you get to above 90% load on the GPU you have found the right setting for that card and more tasks probably will not be more efficient. GPU-Z can also tell you how much of the memory on the card that's used and you can only increase the number of simultaneous tasks until you run out of memory.
RE: RE: There is no
)
Gpu crunching is not officially the goal of either AMD or Nvidia so having a 'manual' at this point is not likely. We have only using our gpu's for a few years now and people are sill experimenting to figure out what works and what doesn't work. In fact at some projects running dual Nvidia cards means Boinc only sees one card, but if you have dual AMD cards at that same project, both work just fine. As Holmis said use gpu-z to determine your gpu's load and follow his guidelines to tweak your system to its best abilities.
There are SOOO many variables that there may never be a real 'manual', but there could be some guidelines someday. I mean what gpu are you using, I am probably using different ones, what about the cpu, again mine is probably different, but each is required to make optimal gpu crunching. How about an Intel or AMD cpu, which is better? How about pc ram, or even gpu ram, neither are standard either. In short SOOOO many variables come into the picture that all help decide whether it is optimal to run 2, 3 or even more units at once on a given gpu or machine. Then they ALL get changed if you add dual, or more, gpu's in the machine.
Then on top of that each project writes its application files differently, making for even more variables. At DistRTgen I was running an AMD 5870 gpu with 1gb of ram and it was doing just fine running 1 unit at a time, I swapped it out and put in an AMD 7970 gpu with 3gb of ram and found that the project itself had optimized the software to run faster the more ram the gpu has. I am still only running 1 unit at a time but instead of 1.5 hours they are no only taking about 22 minutes! The gpu load is identical, the app is just more flexible. The only thing changed was the gpu and the drivers. Obviously not all projects do that, but that one does.
RE: The closest thing would
)
I saw it already via a link from the Russian E@H forum.
GPU-Z, GPU load, etc, I know it for a long time as the IT support engineer. The matter was how don't spend many time in experiments. I had a hope for manual or some tables, more complete than dskagcommunity's. OK, thanks.
RE: RE: There is no
)
Oh, Im sorry, Ive missunderstood your question... I thought you were looking for a way to set each GPU to a different number of simultaneus tasks... My mistake.
The answer you were looking at, is what others have said. And, sadly, there is no way to set a reliable table of performance vs. number of concurrent tasks for each GPU to use it directly.
The maximun number of tasks you can run is limited by the memory usage, that's simple. But what is the right number to get the best performance depends on a lot of different things that are related to the whole host (what model/brand/amount of: CPU, motherboard, Main RAM, OS, number of GPUs installed, etc.), also depends on the number of CPU cores reserved to attend GPU work, and the list goes on...
So the best thing you can get is a guideline like the one already mentioned, that will be usefull only for a certain project, and only while the project applications/versions don't change. But is a good starting point to reduce the number of tests you will need to find your right number.
RE: RE: RE: There is no
)
Another to consider is that not all gpu's with the same number are identical, I have two Nvidia 560Ti's yet one runs units much faster then the other one, the key is even within each gpu things are different. They are different model years so I am guessing some tweaks were made in between each manufacturing date.
RE: Another to consider is
)
Maybe there is not identical GPU really? For example, one of them is the GTX 560 Ti 448 Core?
Other case, both of GPUs have absolutely identical quantity of the ROPs/TMUS/Shaders and identical frequencies, so the cause maybe in the motherboard (the difference in the Bus Interface), for example one card worked at PCI-E x16 mode and other card at PCI-E x4. It can be checked in the GPU-Z or if replace cards mutually.