Einstein GW (O3ASHF search) performance on Radeon VII Pro

[[H]pututu]

@Holdolin , fyi the Einstein GW gpu task credit temporarily got bumped from 5000 to 10000 points per task. I tried this out on my Radeon VII and I could get about 2.15M PPD per card. See here.
Wondering what your Radeon VII Pro with twice the FP64 compute of consumer VII card will perform as there is some advantage with higher FP64 speed. Anyway just keep you informed since Milkyway gpu apps ended and you may be itching to put your Radeon VII Pro to some good use . If you can get higher PPD than Milkyway, that's awesome. The task doesn't sip much gpu power but requires fast CPU for better throughput.

 

[Holdolin]

Sounds great. I just started rebuilding my VII Pro system. Apparently the guide I wrote a few years ago is outdated and no longer works. Need to figure out how to get the software set up and then publish the updated guide. Ironically, in searching the web for this I found a link to my guide, on another site :
https://www.setiusa.us/showthread.p...seeing-AMD-GPU&p=225993&viewfull=1#post225993. It was the only guide I ever published so it was kinda kool to see it helped others. Anyway, I got configurations to figure out and RAC to build.

 

[Holdolin]

Well, this old 2692v3 Xeon is slowing me down methinks. All my fast boards are feeding 4090s so I built this with what I had layin around. Since Bernd has already said he's workin on compiling it for CUDA i'll just crunch away on what I have. If I had a bigger PSU i'd be happy to toss a 2nd card on my board but alas it is what it is. I have to admit I like working with this project much more than some others. While I'm slow to mention names, some projects seem to desire an...adversarial relationship with their donors whereas Einstein seem to foster an atmosphere of cooperation. Also, don't mine me much. It's 7AM and I've yet to go to bed

 

[[H]pututu]

Thanks @Holdolin. It will be interesting to see how the VII Pro outperforms the consumer VII. Yes, you need fast CPU for a portion of the task processing. I'm guessing this might give you the same or higher PPD than Milkyway. Right now, there are very limited BOINC projects where the Radeon VII/Pro card can still close the gap with high end Nvidia cards in terms of PPD output due to its FP64 compute power and high memory bandwidth. I sold one of my VII during the mining boom and now I"ve only one left. How may VII Pro have you got left?

 

[Holdolin]

I got 4, but no fast CPUs to feed them. It seems the days of projects running apps that rely on FP64 power are pretty much over. Glad I noticed this before I started upgraded the VII's to Instincts

 

[[H]pututu]

Well, keep using them until there is no more useful life in it, lol. Don't want them to be a paperweight.

BTW, Einstein have the main BRP7 tasks that replaces the FGRBP1 tasks and the FP64 compute power and high memory bandwidth still helps. If you look at the top 6 Einstein hosts, they all have excellent FP64 compute power and great memory bandwidth. The second best host uses MI100 Instincts cards, I think. The A100 is still freaking expensive but I have the old P100 that used to crunch MilkyWay and it is not that bad for a three generations old card. If and when the GW tasks credit do revert back to the original 5000, there is still a backup project BRP7. My VII does ok in BRP7, about 1.04M PPD the last time I checked but I limit the card power to 150W. Probably can do more.

 

[Holdolin]

So, after a week of massive mugs o' coffee and blistered fingers from exercising my google-foo I figured it out:

### Editing /etc/systemd/system/boinc-client.service.d/override.conf
### Anything between here and the comment below will become the new contents of the file

[Service]
ProtectSystem=off
PrivateTmp=true

### Lines below this comment will be discarded

This <radio edit> (or lack thereof) is why I was unable to crunch in Linux on my AMD cards. Long and short: as is often the case, it was a permissions issue, due to the fact of an improper client configuration that Boinc is quite aware of. I love it when somebody comes on with 'if you need to install AMD drivers in Linux you're doing it wrong', or better 'all you have to do is install OpenCL' Needless to say, I guess I have another guide to write. Though, once you know you need to override some systemd crap the install is pretty straight fwd. Anyway, i need to get the fan control app installed so i can try 4-5 concurrent WUs. However, going from memory, a single WU runnin under Windows was takin ~30mins. The 2 I've completed thus far on my Ubuntu box have completed in ~15 mins. I'll take it....Along with a nice long nap

 

[[H]pututu]

I ran the GW tasks in windows, so I didn'tt encounter any of your issues My consumer VII finished on average 402 seconds (6 mins 42 seconds) per task when running 5 tasks per gpu. Don't quite get it why yours took 30 mins for VII Pro, hmm.

If linux really runs significantly faster, I might consider it.