As some of you know from the other forums, I recently put together a computer from 'spare parts' that has four Nvidia Quadro K2200 GPUs in it. I've been testing various GPU projects to see which ones experience any performance reduction due to the lower bandwidth on two of the PCIe slots in the system.
I am currently testing Einstein@home and noticed that these four cards together appear to be producing about 100-150K credits per day more than my old R9 280X was getting before it had to be retired. I think this is significant because the 280X was rated to pull about 250 watts at full load, but these Quadro cards are only using about 100 watts total (20-25 watts each) at full load.
These cards are too old to run the new apps, so they are not getting higher scores because of that.
I'm not entirely sure since I'm finding conflicting info online, but I believe that these cards do support double-precision computing, which would explain why they get decent results running Einstein. After I let them run Einstein for another day or two to get good averages for my PCIe slot testing, I'll switch them over to MilkyWay to see if they get similar results there as well.
I'd rather have four 3000 series RTX cards, of course, but these were free, and perform surprisingly well considering how little power they use!
One nice thing is that they are very thin single slot cards, so they can fit in boards that don't have space between the PCIe slots (like mine).
I will post a follow-up with the MilkyWay results in a few days.
However, the results are not looking good so far. According to BoincStats, they earned just under 71K total credits from the first day on the project, a bit over 103K credits from the second day, and about 80K additional credits recorded on the project site with around 6 hours to go before the next daily update on the site.
Yuck! They were doing amazingly well on Einstein, but those are pretty pathetic numbers for MilkyWay. Sorry, Skillz...
- TAAT Member
- Posts: 866
- Joined: Sun Sep 15, 2019 3:22 pm
- Location: Massachusetts, USA
Yeah, Einstein doesn't use DP math. Apparently memory bandwith/performance is one factor for Einstein performance but I think it's more complicated than that.
I switched the 4x K2200 GPUs over to Moo! Wrapper as the next project in my testing, and was very surprised to see a single task running, using "0.2C + 4NV (devices 0, 1, 2, 3)"!
That does mean I can't test each card for bandwidth limiting on their ports, but as far as I am aware this is the only project that will run a single task across multiple GPUs. Even though it's not going to give me any info for my bandwidth testing, I think I'll let it run for a while to see what kind of PPD results it gets running all four of them together.
Later it should mainly send work for the best version, but it will periodically send work for the worse ones again to see if they haven't gotten any better. (Flogging a dead horse.) It might even fail to detect the best version anyway.
It is possible to force that only the good application version is run by means off an app_info.xml (anonymous platform). TBD: Determine which application version is the best for a K2200, figure out the app_info.xml contents, which of course also differs a little between Windows and Linux.
One of the worst versions (or the worst version) happens to use all GPUs present in the computer at once, by a single task. Still doesn't help its bad throughput, only complicates matters. BTW, this multi-GPU application version is not really space age technology. All Moo! Wrapper workunits contain several mini-workunits (RC5-72 'blocks'), and AFAIK each of the GPUs simply gets to do a few mini-workunits out of the workunit in this particular application version. The other application versions just perform all mini-workunits serially on a single GPU.
So, not a "280X killer" specifically, but pretty good results in some cases, especially considering the very low power usage on the cards.