Ancestors

Written by Hanno Rein on 2025-01-29 at 02:51

Many thanks to my (former) students @zyrxvo, Dang, Pejvak and Nick, who helped disassemble 16 nodes of the NIAGARA supercomputer today. Everything is now at UTSC and once we've put it back together, my group will have its own small compute cluster with 640 cores.

=> View attached media | View attached media | View attached media | View attached media

=> More informations about this toot | More toots from hannorein@mastodon.social

Written by Oliver Stueker on 2025-01-29 at 12:38

@hannorein

Really Cool!

It’s nice to see that these Niagara nodes get a second life while making space for Trillium!

However now just imagine that Trillium will have 192 core per node and therefore you could cram these 640 cores in 3 1/3 nodes. 🤯

[#]HPC #SciNet #Alliance #DRAC

@zyrxvo

=> More informations about this toot | More toots from ostueker@mast.hpc.social

Toot

Written by Hanno Rein on 2025-01-29 at 12:45

@ostueker @zyrxvo Yeah, it'll be interesting to compare the performance per core with my specific code. (I suspect not much change - despite the 7 year time between the two).

Small downside of the new Trillium cluster: I will probably never be able to give those a second life because everything is watercooled...

=> More informations about this toot | More toots from hannorein@mastodon.social

Descendants

Written by Oliver Stueker on 2025-01-29 at 23:58

@hannorein

No, I don’t expect a huge difference in per-core performance either, though I haven’t compared benchmarks yet. Niagara's Skylake CPUs already supported #AXV512 (they even waited a bit for Skylake to be released) which gave a good #performance boost over AVX2. But core-density is the new MHz. 😉

Though you may want to recompile your code if you’re currently directly linking to #MKL since Trillium has AMD CPUs. BLIS is faster on those.

[#]HPC @zyrxvo

=> More informations about this toot | More toots from ostueker@mast.hpc.social