Glenn K. Lockwood
@glennklockwood
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.
This is the diagram I’m most proud of. Stacking ~700 different LLM training jobs observed from the VAST fleet on a common timescale to figure out whether there is a “feeding the GPU” problem anywhere close to the write intensity of periodic checkpoints. There is not. Only read burst is at job start.
Nobody came 😩 Just kidding. Was mic check. Here’s the actual crowd.
We are now in the age of AI-generated conferences. Today’s keynote by Gropp prominently features distinctly Claude-generated plots; yesterday had an entire presentation delivered by an AI voiceover of slides. And slides are full of Claude design tells.
First time I’ve seen someone with a PhD actually predict that data centers in space will be a real thing. Granted, it’s a CS PhD, not a physical sciences one. #HPC #notreallythough
Interesting that C-DAC, whose indigenous Indian interconnect Trinetra is based around libfabric, opted to implement verbs support to make Lustre work instead of adopting Cray’s kfabric LND. #ISC26
1. General LLMs beat coding-specific models at writing Fortran, likely bc code models are overfitted to Python 2. This conf begins the age of Claude-generated pptx. Clear tells 3. My phone doesn’t like these #ISC26 projectors
He has a chance to mention the Ozaki scheme but instead cites a 20-year-old algorithm that is the problematic core of HPL-MxP.
I wonder if ground has been broken on this datacenter yet. Because it takes two years to build such a facility without the federal land red tape part. How commercially meaningful will Blackwells be in two years?
Funny that “software adapt to whatever hardware shows up” is exactly how everyone’s favorite AI models are developed. Model architecture iterates much faster than chip design. This only works if your problems stand still.
How many wrong statements can you fit into a single sentence? Holy cow.
It’s comforting to know that, at the macro scale, Lustre is pretty consistent at its strengths and weaknesses. I don’t think these are different from when I last used Lustre in 2022. From @adrianjhpc.bsky.social at the Lustre BOF in #ISC26
Supercomputer for agents is a bold bet by SJTU and a great thought exercise (if not more). There’s a lot of crap that we drag along in the #HPC stack simply because it makes humans comfortable.
I got to work with VAST's architects on integrating NVIDIA Vera BlueField-4 STX in-silicon security features with VAST's agent policy enforcement engine. Agent interactions can be inspected and governed at line rate, and policies can adapt in real time. Very slick. www.vastdata.com/blog/vast-ze...
the text labels were mine; I had to obfuscate what I was actually trying to do (invade Europe). but the chorems are fascinating! looks like something else that Claude recommended I do to reduce the size of my svg of Europe.
Claude offered to create an SVG of Europe for me (I did not ask), and this is what came out. We have a long ways to go before the Terminators take over.
Found my new LinkedIn profile while visiting my mom this weekend
A career in collaborative grant writing, writing papers, and peer review has prepared me for spending days (weeks?) of scrubbing deep technical marketing copy.
I just got confirmation that my login credentials on NERSC supercomputers were revoked at some point and no longer work. With my Microsoft free Azure gravy train also ending late last year, I am officially without a supercomputer. First time in my career 😢
The shiny silver rack door for NERSC's upcoming Doudna is quite different from the Cray racks that NERSC historically had. Wonder how the rack graphics will work. End-of-row, like OLCF Summit's? (also shown: an 11-node VAST cluster in the second rack) #HPC Source: www.linkedin.com/posts/michae...
No wonder there is such vast unmet demand for tokens. Filling out one government form has burned up two-thirds of my $20/mo token budget in fifteen minutes.