As a long-term but long-retired (2004) Intel employee and longtime Einstein participant, I'm well aware that Intel has not often been a good bet for graphics, nor a much-desired Einstein host component, but I wonder whether the Arc B-series cards being introduced in December 2024 might find a place.
At first glance the B580 seems to be aimed at a place in the price/power/performance spectrum that is a bit ill-served by current offerings.
If the drivers let it actually work with Einstein, then the question would be whether the hoped-for "modern" price performance and power performance might find it a place for people not wanting to play at the high end of the game.
To be more specific, I'm typing on a PC currently running an Nvidia RTX 3060 card. I'd be happy to try out one of the new Arc B580 cards if there is a decent chance it might work, and might actually compete favorably with the RTX 3060 on price performance and power performance.
I'm starting this thread, hoping people will chip in with evidence for and against the possibility that this might be a pipe dream on my part.
Copyright © 2024 Einstein@Home. All rights reserved.
We would be willing to get in
)
We would be willing to get in on this too- we are all about education and learning/trying new things. When they do come out, we might be able to pick one up and put it through the paces.
They can run opencl, so I would think they could work?
Einstein and some other Boinc
)
Einstein and some other Boinc projects already have OpenCL apps for Intel ARC GPUs, the next Battlemage generation should work just fine. B580 cards should compete with 4060 while being cheaper, so it should be faster than your 3060.
However, power efficiency has not been a strong point for Intel GPUs. While Battlemage may close the gap with the current generation of competitors, official benchmarks have yet to be released...
I'm trying to run on the B580
)
I'm trying to run on the B580 but not getting any work. Maybe there's a filter on the server side that doesn't send work to BMG yet?
Haven't followed the types of workunits available for quite a while but my other hosts are doing BRP7. I have BRP4G, BRP7 and FGRPB1G enabled.
Mumak wrote: I'm trying to
)
Try reading this thread, Keith Myers helped him get his up and crunching:
https://einsteinathome.org/content/getting-tasks-intel-hd-graphics-5500
Thanks! 1st BMG BRP7 WU done
)
Thanks! 1st BMG BRP7 WU done on E@H - 563 sec
that's decent performance
)
that's decent performance actually. how much power is it using?
_________________________________________________________________________
I don't have the Powenetics
)
I don't have the Powenetics PMD hardware fully assembled yet (need to connect the PCIe riser) so can't verify precisely.
But according to software monitoring (which should be quite accurate now) the average TBP (Total Board Power) is 111 W.
Currently the PMD-measured PCI connector avg power is 65W, so given the slot can deliver max 75W, it will be < 140W. Will give more precise data tomorrow.
software measurement is good
)
software measurement is good enough. that's not bad especially from an OpenCL app in Windows. probably not the best choice for Einstein or other BOINC projects (most projects favor Nvidia) but having hardware FP64 capability now is a good thing.
I know it's an older card, but as a point of comparison, a Titan V can be had second hand for about $300 USD, and will do a BRP7 task in ~200s (Linux, custom application, software tuning) using ~135W.
_________________________________________________________________________
Hmm, Titan V @ 135W, that
)
Hmm, Titan V @ 135W, that must be with heavy optimizations. Wondering how it was measured.
I'm pretty sure BMG drivers will be improved, this isn't running final (retail) version yet.
Did a quick PMD verification of true power (slot+connector) and it was 112 W. Incredible software accuracy.
measured via software
)
measured via software reporting just like the Arc uses.
yes, heavy optimizations.
_________________________________________________________________________