iiinova — where your digital self keeps evolving

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 21/10/2026

Memory, storage, and GPUs are so expensive in part because hyperscale GPU clusters are only 87% usable. The remaining 13% are kept as powered-on hot spares that sit idle. Could you imagine #HPC doing this? […] [Original post on mast.hpc.social]

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/10/2026

TIL that essentially NONE of the AI cloud providers are exhibiting at #SC26, though many left in 2025. Seems the overlap between the SC #HPC community and hyperscale #AI infra industry is asymptotically approaching zero. There's also a ton of churn in the […] [Original post on mast.hpc.social]

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 07/10/2026

@AlanSill That opener was cute in 2023, but anyone using it in 2026 is out of touch.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/09/2026

I spent the last few months preparing the technical docs for DataEnclave, VAST's new confidential computing runtime. I had to learn tons about how attestation, TEEs, and related technologies REALLY work. VAST's R&D wizards made it possible for me to learn it to the point where I could describe […]

mast.hpc.social

Original post on mast.hpc.social

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 22/09/2026

By total coincidence, two of my colleagues are running a webinar on this topic on Wednesday. The lead speaker, Anat, is really sharp and worth hearing speak […]

mast.hpc.social

Original post on mast.hpc.social

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 21/09/2026

I wrote this doc to explain KV caches and KV cache offloading many months ago. I found salespeople couldn't confidently tell when their customers might have a problem, because all the content online is straight marketing or way too in the weeds […]

mast.hpc.social

Original post on mast.hpc.social

110 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 17/09/2026

Malaysia is allegedly evaluating Huawei's Ascend 910C for their sovereign AI buildout. If Malaysia picks Huawei over NVIDIA, it'd be the first big buy of these chips outside of China. Egypt also name-dropped as a nation being pursued by Huawei […]

mast.hpc.social

Original post on mast.hpc.social

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 10/09/2026

The IO500 committee just stripped Sugon of its #1 position in the Production list because it appears that their ParaStor F9000 isn't a real product, or it could not be verified as one. Verify here: io500.org/list/isc26/production My notes on ParaStor […]

mast.hpc.social

Original post on mast.hpc.social

020 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 03/09/2026

I recently had the opportunity talk to my pal Nicole about a topic I presented at PEARC'26 this year: failure. We chatted about a few failures in #HPC and #AI infrastructure that I was a part of but were hidden behind headlines of success across my time at SDSC, NERSC, and Microsoft. I spoke of […]

mast.hpc.social

Original post on mast.hpc.social

001 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 01/09/2026

MLPerf Storage results are out and StorageReview pointed out the interesting disconnect between MLPerf leaderboard winners and companies with the largest #storage footprints in the #AI clouds. Suggests this benchmark may not reflect real needs […]

mast.hpc.social

Original post on mast.hpc.social

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 22/08/2026

I recently wrote up an argument about why batch schedulers like Slurm are architecturally insufficient for the needs of autonomous scientific laboratories and agent-driven experiments. Its serial scheduling algo is great for maximizing utilization, but it collapses when hundreds or thousands of […]

mast.hpc.social

Original post on mast.hpc.social

011 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/08/2026

Funny, there’s a reason NERSC took its datacenter out of this location. It’s not a great place to put one. www.datacenterdynamics.com/en/news/…

020 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/08/2026

I have a theory that every panel about “how AI will affect XYZ” will devolve into nonspecific discussion around “here is where the AI tools I use today fail” and “how do we trust AI?” regardless of what XYZ is.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/08/2026

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.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/08/2026

The lack of shame some people have in copypasting prose straight out of Claude into Powerpoint is shocking. Were these people just recycling their students’ slides before, and now just tap a different source of material?

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 10/08/2026

Problem is, it's not my day job to do this kind of analysis. So here I am, frantically banging on signal analysis and statistical models on a Sunday, to have something compelling to say 😩 As with all longitudinal workload analysis, people are messy and hard to understand or predict.

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 10/08/2026

I am attending the ModSim conference for the first time next week, and I'm presenting analysis and design modeling of I/O performance requirements for frontier model training based on a year's worth of production data across the VAST hyperscale fleet (>900 training jobs, >140K GPU-years of compute)

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 07/08/2026

I recently upgraded to Claude Max and started using Opus 5 for exploratory analysis. If this is what the AI community things approaches "PhD-level capability," I don't know what to say. It's sloppy, inconsistent, and walks back conclusions at the lightest of questioning.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 01/08/2026

RE: mstdn.ca/@dbattistella/117004005387… Where is the evidence that rare books were destroyed? This is a fun trope to hate on, but it doesn’t make sense; these books were purchased in bulk (source […]

mast.hpc.social

Original post on mast.hpc.social

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 22/07/2026

If you hate reading HPC marketing material, imagine having to write it

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 15/07/2026

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.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 14/07/2026

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

110 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 08/07/2026

@jannem Yeah that’s fair. My perspective is colored by my own history in Top10 computing with the US DOE.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 02/07/2026

I somehow missed that China has developed its own implementation of InfiniBand, called ScaleFabric 400, to compete/supplant NVIDIA/Mellanox IB. Oddly enough, its manufacturer (Sugon) is not a member of IBTA though. en.eeworld.com.cn/mp/EEWorld/a42265… #HPC

002 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 26/06/2026

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

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 26/06/2026

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

101 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 25/06/2026

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.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 25/06/2026

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?

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 25/06/2026

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.

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 25/06/2026

How many wrong statements can you fit into a single sentence? Holy cow.

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 25/06/2026

I can’t wait to hear what Jack Dongarra thinks about the future of #HPC at #ISC26’s closing keynote.

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 24/06/2026

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.

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

Hot #ISC26 tip: VAST is running free taxi service from the taxi stand in front of CCH to the SIDE hotel downtown and back. If you don’t feel like getting sweaty before/after an offsite meeting, I think anyone can hop in. Look for the city cabs with the VAST logo on the side.

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

The jump of LineShine over El Cap in efficiency is staggering. Suggests the benefit of balance and design for FP64-first over wedging FP64 into a socket whose chiplets were AI-first. LineShine’s poor HPL-MxP underscores. #ISC26 #HPC

021 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

“Heterogeneity is here to stay.” Also, “the new #1 system is homogeneous all-CPU.” #ISC26 #HPC

110 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

China’s back at the top of Top500. I guess the threat of not being allowed to buy US-designed chips is no longer a compelling reason not to submit top systems. Happy 250th, USA. #HPC #ISC26

012 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

@jannem Was it restraint, or just unabashed “I am not working on AI and that’s OK”? It was a true HPC keynote for an HPC audience

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 23/06/2026

20 minutes into the opening keynote at #ISC26 and Martin Schulz has managed to say AI only once; it’s changing everything, but let’s talk about quantum and #HPC instead. Impressive.

100 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 13/06/2026

The sense of entitlement from the highly paid employees at these big tech companies is flooring. If these companies treat you so badly, why don’t you just work at a better company? It’s what I did. There’s no free lunch in life […]

mast.hpc.social

Original post on mast.hpc.social

010 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 11/06/2026

@kpdooty These guys? They’re the official bird of Oakland! I can send you some if you want; they’re always in our garbage.

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 11/06/2026

"DDN is seeking to attract investors that can help the company better reflect the value it creates" One of the weirder statements I've seen recently. Implies that current investors do not value their investment as highly as the investment values itself. Maybe par for the course for #AI […]

mast.hpc.social

Original post on mast.hpc.social

001 1

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 10/06/2026

Was reminded that PEARC26 regular registration closes this Friday! It's a great, practitioner-focused #HPC, #RCD, and #RSE conference, and I will be there this year. pearc.acm.org/pearc26/registration

021 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 09/06/2026

RE: mast.hpc.social/@AlanSill/116718896… "FP8 is all you need" vs "A naive Ozaki II implementation with r=10 residue planes per warp can spill, raising β above 1 and erasing the projected speedup." Which one wins? #HPC

000 0

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 08/06/2026

I am convinced there is a growing lexicon of AI terms that are used exclusively by people who don't understand AI to talk to other people who don't understand AI. For example, has any real AI practitioner ever referred to "keeping GPUs fed?" #AI

021 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 03/06/2026

For my #HPC friends, it also describes the stack used to run bulk-synchronous training jobs entirely without Slurm: k8s, kueue, and Ray Train. Superior at scale because a single GPU failure doesn't crash the job; a new worker pod is spun up, Ray registers the worker, and restart is lightweight.

021 2

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 03/06/2026

Microsoft AI's report on MAI-1 training is full of interesting infrastructure anecdotes, especially around storage: - Async checkpoint from DRAM to object (no node-local SSD used) - Restart is read once from object, then broadcast over InfiniBand to avoid incast - Anyscale Ray for fast restart […]

mast.hpc.social

Original post on mast.hpc.social

132 3

Glenn K. Lockwood @glennklockwood.mast.hpc.social.ap.brid.gy · 03/06/2026

Virtual DAOS User Group slides just dropped (daos.io/event/virtual-dug26). Interesting content, and glad DAOS is still trucking along. Surprises: ALCF has only 37% servers in production; LinkedIn trying to run DAOS on 20% of cores on HGX servers; long feature roadmap. #HPC #storage

000 0