Profile avatar
glennklockwood.com
I am a supercomputing enthusiast, but I usually don't know what I'm talking about. I post about large-scale infrastructure for #HPC and #AI. Disclosures: Employed by Microsoft. I used to work at NERSC/LBNL.
808 posts 1,044 followers 196 following
Regular Contributor
Active Commenter

At ISC last week, I noticed that a LOT of sovereign AI systems are being designed more like traditional modsim rather than pure AI (per this doc). Wonder if those sovereign AI systems are really going to be used for as much AI as advertised.

The only compensation I get for serving on the SC technical program committee is the right to complain about it. So here we are. At 6am. 😫

Great news for AMD, but boy is this press release confusing: “OCI to deploy new zettascale AI cluster with up to 131,072 MI355X GPUs” So are they or aren’t they actually building a cluster? And 131K MI355X only hits zettaflops at FP4, which isn’t useful for training. www.oracle.com/news/announc...

Looking at the Top500 list from #ISC25, it looks like JUPITER ran HPL with 5,884 nodes--perhaps the full system (advertised as "roughly 6,000" nodes).The run only hit ~33 TF/H200 tho; by comparison, Isambard-AI hit 43 TF/H200. If JUPITER can get up to 43 TF/GPU, it should post at 1.01 EF FP64. #HPC

It's official: #ISC25 was the largest iteration yet, with 3,585 attendees registered. It didn't _feel_ like the biggest one yet, but that's not a bad thing. I had a blast--genuinely the most fun I've had at work all year. There was a lot to chew on this year. isc-hpc.com/isc-2025-con...

Just because the “infinite workday” has been quantified doesn’t mean it’s new. Split shift existed before the pandemic and even before smartphones. It’s just easier to be more productive off-hours now than ever before. www.axios.com/2025/06/17/m...

The "agentic mesh" is like an office; you have to build it before workers can get together, and it may take months/years for the benefits to manifest as workers figure out how to collaborate. It's a longer-term investment.

An important, but often overlooked, point in the AI discussion is that AI always acts under the delegated authority of a human. Call them sin eaters or whatever, but AI will just be yet another entity following orders in a chain of command.

Our reflections from #ISC25 are live. From standout sessions to big-picture trends and community moments (including our Roco casting a few spells), we’ve pulled together the key takeaways from Hamburg in our latest blog. 📬 Read here: buff.ly/Lnmoyfd #ISC#HPC

Catch up on ISC gossip and talk next generation schedulers…. Oh and your usual weekly #HPC news cloudhpc.news/quantum-64/

@micheleweiland.bsky.social talks about sustainable supercomputing @epcc.bsky.social , including clock frequency on #ARCHER2, malleable jobs, novel hardware and teaching. At #PASC25, more details on our @cerebrassystems.bsky.social systems tomorrow afternoon by Justs as part of #CONTINENTS

Jake Davies, one of my PhD students @epcc.bsky.social presents our work on accererating #FFT on @tenstorrent.bsky.social #wormhole at @riscv.org.web.brid.gy for #HPC . #wormhole is around three times more energy efficient than a 24-core Xeon Platinum and I think we can push this even further

I had no idea Fortran for deep learning was anything but a joke, but this talk from LBL is interesting. Apparently it can be really fast (cf. pytorch), and many modern Fortran features may map nicely to the algs used in ML/DL. #ISC25

Nice work from U Bologna: Tenstorrent has a good, if a bit complicated, story around efficiency (performance and perf/watt. What’s not shown in these slides is the massive price difference. At $699, I kinda want to buy one of these accelerators to mess with. #ISC25

Easy misconception is that AI people can just downcast from high precision to low precision and life is good. Reality is that there’s a bunch of gnarly mixed precision going on that’s a lot of work and complexity to implement usefully. Nobody’s enjoying a free lunch in either HPC or AI. #ISC25

Last day of #ISC25 is often the best day! I’m presenting at SuperCompCloud in Hall 10 at 1pm on how we effectively (and sometimes ineffectively) apply hybrid HPC inside Microsoft.

Interesting to see Yutong Lu cite Top500 to build her case without acknowledging any of the Chinese exascale systems. She invoked the China Top100 list in a talk the other day, but it’s much less detailed. #ISC25

According to Liran Zvibel, “experience is what you get when you don’t get what you want.” I like it. He also couldn’t help but proclaim WEKA as the best parallel file system ever while on stage. I guess that’s what sponsorship entitles you to. #ISC25

Fun presenting with @fclc.bsky.social at #ISC this week, at the @riscv.org.web.brid.gy BoF and then today at the democratising #AI accelerators for #HPC . #RISCV continues with a workshop Friday afternoon, kicked off with keynote from @riscv.org.web.brid.gy CEO. riscv.epcc.ed.ac.uk/community/work…

I am getting tired of hearing about FP64 emulation. But then I realize I only first heard about this ~9 months ago. Ozaki scheme has a hell of a marketing team. #ISC25

I’m not sure these opportunities are super compelling. If your accelerator is so novel that it requires CS research to use, it’s probably not actually solving domain problems. It’s more like CS for CS’s sake. #ISC25

At the panel on energy efficiency beyond hardware, the topic of “why can’t we do what cell phones do?” came up. Amazed that #HPC people don’t understand that the latency requirements of a human-driven device are way easier than RDMA and memcpy. #ISC25

The AWS party at #ISC25 looked like it was the place to be tonight. Sadly, a bunch of us Azure folks forgot to RSVP and were turned away at the door. Maybe next year. However, the later party sponsored by HPE was fantastic! Beautiful venue and lots of great conversation. 5/7 would recommend.

Happened up Dan Stanzione showing TACC’s results from the Grace CPU vs SPR. Surprisingly not compelling in the absolute sense, but the perf/watt is where it shines. Surprised NVIDIA lets Dan talk facts rather than NVIDIA’s usual 1000x speedups with unqualified plots.

Sounds like KISTI is getting a new 600 PF GH200 Cray EX as its KISTI-6 system. Wonder how long before you need >500 PF to make it to Top10. #HPC #ISC25

#AI is the biggest chance and opportunity that our community has had since @thoefler.bsky.social can remember! Our field is growing! The community is flourishing! AI is positively shaking the community and pushing it forward!! #HPC #ISC25 Fishbowl Panel

I guess I’ll have to retire this pin now.