Posts by WMD

1) (Message 7539)
Posted 4 Dec 2022 by WMD
Hi Georgi,

I remember from 2021 and earlier, the project would run out of work units fairly frequently, because you weren't able to get new units fast enough before the current set of units were complete... and also, it was partly caused by people hoarding large numbers of them at a time. Do you expect the same pattern going forward? (I'm not complaining, I just want to know what to expect so I can set up my BOINC machine accordingly.)
2) (Message 7366)
Posted 19 Nov 2022 by WMD
that GPU can do a task 10x faster, at the same power, but the CPU can run 32 of them in parallel. making the GPU slower overall.

If you have a CPU with that many threads. A lot of people don't. (Or, you may not run this project with all available threads/cores.)

also the apps here seem to benefit a lot from FP64 performance. since the project is Nvidia only for the GPU apps, it's hard to see that benefit since most consumer nvidia GPUs have lackluster FP64 performance. I'll have a TitanV here soon to verify that.

I have a Titan V, it seems to average between 6.5 and 7.5 minutes per work unit. (I only run it as a fill-in project, so I only do a handful of units at a time.) Makes sense that FP64 would be advantageous here, since the numbers involved are, dare I say it, astronomical. On the other hand, that speed comes out to about 8.5 work units per hour, and it seems a lot of CPUs can do 1 unit per hour per thread, so, quite a few CPUs would be able to outperform a GPU.

Personally, I try to run CPU-only projects on the CPU, and GPU on GPU. This project, as I mentioned, is an anomaly for me.
3) (Message 7338)
Posted 17 Nov 2022 by WMD
This is amazing!!!

I had left the project (and two unreported tasks) open in my BOINC setup, but was losing hope that I'd ever be able to submit them. Sure enough, today I noticed they were gone and a new one was being worked on! (This has always been a secondary project for me, to fill in the gaps where MilkyWay@Home doesn't send work. So it took a few days for me to notice.)

I checked my task status, it looks like the ones I had in my queue for over a year made it in and were even credited. :D

Happy to be crunching again!
4) (Message 7337)
Posted 17 Nov 2022 by WMD
With a GPU you can try other projects. Asteroids@home GPU tasks are not significantly faster than CPU tasks. As a result, those top computers are not equipped with a high-end graphics card.

That statement is patently false. GPU apps are 2X-10X faster than any cpu at this project.

Indeed... my GPU can finish a unit in 6-7 minutes, but I see CPU hosts taking an hour or more. I think the top hosts being CPU hosts is because they have crazy CPUs - #1 is a dual-Epyc box with 256 total threads!

So only the Turing and earlier cards work here. You won't get sent work for any Ampere cards.

My Volta card works too. (I can't remember which is newer, Turing or Volta...)

Either way, the code should definitely be updated for Ampere. There would be a nice performance bump there!