Well... I submitted the same message in forum at Einstein@Home. And I may have to try it here as well:
Recently we got box of nVidia's DGX Spark GB10 (yes... that one what was claimed to be AI supercomputer on desktop) and I tried Minecraft@Home BOINC client on top of that since they have aarch64 + cuda application developed and deployed.
Well... it's quite interesting (certainly it's a little ridiculars to use it with BOINC via a box costs US$4000+) to try it by SRBase.
In that case, maybe spare some time for development team to build (compile?) a variant application with aarch64 + cuda/opencl if that does not bother too much effort on porting source codes?
Thanks a lot
/ 
Reply
Quote
I think the existing TF CUDA 12.9 build for Jetson devices should work, but the server isn’t accepting it (no work available)
https://srbase.my-firewall.org/sr5/forum_thread.php?id=1946
EaH O4AS tasks take about 800s (with a custom optimized app)
That said, in general, the GB10 is a terrible choice for boinc compute workloads once you factor in cost. Yes, 128GB of VRAM is cool, but it's only useful for AI.
/ Reply
Quote
I think the existing TF CUDA 12.9 build for Jetson devices should work, but the server isn’t accepting it (no work available)
https://srbase.my-firewall.org/sr5/forum_thread.php?id=1946
EaH O4AS tasks take about 800s (with a custom optimized app)
That said, in general, the GB10 is a terrible choice for boinc compute workloads once you factor in cost. Yes, 128GB of VRAM is cool, but it's only useful for AI.
How would we make the VRAM more useful?
____________
Primes Found:
50*893^231310-1 (Top 5000)
20*634^476756-1 (Top 5000)
96*789^506568+1 (Top 5000)
225*370^322863-1 (Top 5000)
3293*82^335630-1 (Top 5000)
18431*82^467690-1 (Top 5000)
79128*105^365950-1 (Top 5000)
More primes on prestacki.com!
/ Reply
Quote
Math projects don't require much VRAM (typically < 1GB)
The most VRAM hungry Einstein O4MDG tasks require about 7GB/tasks, running more of them could help server cards like B200, but they also have good bandwidth (HBM)
GB10 has plenty of VRAM, but its memory bandwidth is only 273.2 GB/s, so running multiple tasks won't scale well.
Even ignoring the price, I can't imagine any projects where this ratio would be useful. Also, building a project that requires more than 32 GB of VRAM would limit the volunteer pool to just one or two people with access to such hardware. Bad idea...
/ Reply
Quote