New SC record: ran into a colleague within 2 minutes of walking into the airport terminal from the curb. Been catching up nonstop straight through takeoff. Conference starts earlier and earlier every year.

Chatting with a pal reminded me of a fun pre-SC activity: looking back at old conference takes that aged like milk. Remember this one? #hpc #zettascale #hedoesntworkthereanymore

🎶 It's the most wonderful time of the year 🎶 #sc25

VAST and CoreWeave just announced a >$1.1 billion partnership to deliver #ai data services. Mind you, that's a billion in services, not GPUs. Though I can't claim any credit, I'm proud to work for a company that's earned this level of trust from a partner […]

SC25 will be my 12th SC (10th in-person). I've attended and presented on behalf of SDSC, NERSC, Microsoft before, but I've got to say: this year has been the most work and most stress I've ever had around the conference.

Google recently posted a promo for using their managed #lustre service to accelerate inferencing via KV caching. Raises questions: 1. What ever happened to Google Managed #daos (ParallelStore)? It performs better than Lustre. 2. Does Gemini use this? Unlikely. See […]

This is like showing up with a new boyfriend the week after the divorce. At least Microsoft is still getting those alimony payments. https://www.aboutamazon.com/news/aws/aws-open-ai-workloads-compute-infrastructure

NVIDIA, Oracle, and US DOE are named in the headline. Argonne is not. I don’t think this is an Argonne system. https://nvidianews.nvidia.com/news/nvidia-oracle-us-department-of-energy-ai-supercomputer-scientific-discovery

"AMD Powers U.S. Sovereign AI Factory Supercomputers" - What exactly is "US sovereign AI?" I think the whole point of "sovereign AI" is "not dependent on the USA." All AI is, by default, sovereign to the US […]

Few more details on ATS-5/Mission #hpc to be installed at LANL. Confirmed as next-gen Cray GX5000 with Vera Rubin + XDR InfiniBand. Confirms that GX5000 can do both InfiniBand and Slingshot++. The messaging is funny; it "will build on the success" of LANL's Venado (non-ATS; GH200) system" but […]

Kinda funny that OpenAI owns less of OpenAI than Microsoft. OpenAI (the nonprofit) holds a $130B equity stake in OpenAI (the public benefit corporation) while Microsoft holds $135B. https://openai.com/index/built-to-benefit-everyone/

NVIDIA announced seven new DOE #hpc systems for #ai, including a 100K Blackwell system at Argonne. Oracle is a partner, just like with yesterday's OLCF-6/Discovery announcement. Details on the ALCF systems are scant. This procurement was out of band of the Aurora follow-on. The LANL systems […]

Article implies AMD is partly paying DOE for the #hpc systems being deployed at ORNL: “The Department of Energy will host the computers, the companies will provide the machines and capital spending” Is AMD buying market share, similar to its arrangement with OpenAI? […]

Glad to see DAOS getting another round of DOE investment. Highlights a diverging gap between government AI infra and commercial AI infra. I have never met an AI customer who's found relevance in DAOS. Maybe .gov can change that […]

I published my first (of many) technical blogs for VAST. This one gives a quantitative, real-world perspective on how much checkpoint bandwidth is required to train trillion-parameter-scale models (hint: less than many have suggested) […]

KV caching for LLM training is just a specialized memoization problem which is not uncommon across #hpc, and its implementation remains pretty primitive but is optimizing quickly. Here's an example of scavenging "perforated" results of memoized attention layers […]

Love to see DAOS getting mainstream attention, but it's long road ahead before it can compete in the GPU cloud market. Everyone says their performance is the best; that doesn't differentiate storage for #ai. It's the hard problems - integration, reliability, etc […]

I'll be talking about how hyperscale #ai resembles #hpc workflows at an #sc25 Exhibitor Forum session on Thurs 11/20. Excited to be on stage at SC again. Aiming to maintain the same level of technical quality as before I left .gov. Details […]

Something eye-opening I've learned since leaving MSFT/joining VAST: how many companies out there just want to turn space+power into $$$ via GPUs. They have zero interest in #ai or technology; it's just the next BTC. This partnership is for them. They can get a turn-key GPU datacenter with next […]

Although I am not officially attending OCP this year, I will be participating in an early morning breakfast panel that will discuss future directions of #ai infrastructure. If you'll be in San Jose that week, consider attending! https://solidigm.techarena.ai/ocp-panel/ #shameless #selfpromotion

I’m in North Dakota this week celebrating Scandinavian heritage. That liquid on top is butter.

Amazes me that people can claim to be an expert, regurgitate what others have said, and have audiences actually believe they are visionary. Do they truly drink their own kool aid, or do they know that they are actors on a stage? Genuinely curious what it takes to be that kind of person.

Visited some of my old pals at NERSC this afternoon. It feels like a lot has changed, but at the same time, it’s the same old place. I do miss working here, but am glad I now get a chance to visit under the auspices of my new role at VAST.

Only took 2.5 years to implement quote posts (see https://mastodon.social/@Mastodon/110294411952997299) From: @Techmeme https://techhub.social/@Techmeme/115192729559586404

Facts aside, you know marketing is effective when competitors feel the need to engage and write their own opinions on it. Where it started: https://www.solidigm.com/products/technology/economics-of-exabyte-data-storage.html Where it's gone […]

I expect https://arxiv.org/html/2508.15884v1 to be a hot GTC topic; NVIDIA is claiming credit for >50x increase in inferencing throughput based on algorithmic improvements. Lets you either (1) reduce KV cache requirements (good for local models on phones) or (2) dramatically increase context […]

New paper on power swings caused by bulk-synchronous workloads published by my old team at Microsoft + OpenAI (https://arxiv.org/pdf/2508.14318). Some flexing ("the utility company requires us to meet a 10MW dynamic range, and we are running a job using 100MW of power") but good workload insight […]

Though it all kinda looks the same these days, I'm a sucker for photos of #ai #infrastructure, and Patrick at STH just posted a great walk-through of a DGX B200-like supercomputer that Lambda Labs has in Ohio. Even has some photos of VAST storage! […]

ACHIEVE, a 14-partner EU project led by EIT Digital, is gathering industry input to align #hpc, Cloud & Networking training with real-world needs. Here’s the survey […]

Getting on a plane for VAST for the first time on Monday. But I’m not so sure about the destination 🥵

GPT-5 is cool and all, but it's still not to be trusted. Didn't take me long to get "PhD-level" made-up facts. What really distinguishes "PhD-level" from whatever GPT-4 was is the ability to critically think, be skeptical, and stand behind the information […] [Original post on mast.hpc.social]

One thing I really enjoy about working at VAST (or perhaps that I enjoy about not working at Microsoft) is that I can go out and talk to people again as part of my job. Here’s a view from where I got to spend my afternoon today. #nothpc #butthatsok

I picked a good time to quit, eh? I still have the value of my unvested stock in my Yahoo Finance, so I get a daily reminder of how much I walked away from 🙂 Satya’s no dummy, but I do worry that the blowback from the heavy layoffs during blue skies was underestimated. Cutting people to drive […]

Very cool to see new neuromorphic coming online. Excited to see where this architecture goes beyond TrueNorth. “4320 chips with 152 cores each. The chips are 48 to a board with a single chip consuming between 0.8-2.5W. The whole system fits in a single rack” […]

The 13th MVAPICH User Group (MUG) Conference is coming up (Aug 18-20 in Columbus). It's a great event for focused technical presentations around MPI and network performance. Great speaker lineup too. Wish I could make it in-person, but virtual attendance is free […]

I wrote up some notes on how to approach I/O and storage benchmarks in RFPs. I normally don't post here about updates to my digital garden, but I think this page is tidy and useful. https://www.glennklockwood.com/garden/benchmarking#storage #hpc

TIL that the Linux NFS client recently accepted a "noalignwrite" option that allows you to safely do shared-file writes to non-overlapping, non-4K-aligned extents. No longer have to use direct I/O to do shared-file writes to NFS with this enabled. See […]

I am a sucker for photos of cool #HPC infrastructure, and here is a dense GB200 NVL72 cluster going up somewhere in Canada (I think). Impressive to see this many racks in row; the DC must have facility water which is still uncommon in hyperscale. Source […] [Original post on mast.hpc.social]

Upon hearing I left MSFT, founder of a big GPU cloud provider reached out with a very strong offer to join them. Funny thing: I applied there back in April and was auto-rejected. Moral of the story: don't chuck an application over the wall if you have an inside line! Applying online is a crapshoot.

In the few days I have between jobs, I wanted to share an unvarnished perspective on what I've learned after spending three years working on supercomputing in the cloud. It's hastily written and lightly edited, but I hope others find it interesting […]

Today is my last day at Microsoft. I’ve learned a lot over the last three years, but I’m ready to try something different.

In the last 18 hours, I’ve learned way more about adjuvanted vaccines and adverse reactions to them in cats than I ever cared to. And the real kick is that the choice to re-up the cat’s vaccines was an afterthought, because we had to take other cat in for […] [Original post on mast.hpc.social]

This is cool, but the real proof is in the quality of the frontier models that are trained on Blackwell. And by that metric, GB200 NVL72 has yet to deliver anything. https://www.coreweave.com/blog/coreweave-leads-the-way-with-first-nvidia-gb300-nvl72-deployment

NERSC just announced that IBM and VAST have been selected as the storage providers for the upcoming Doudna #hpc system. Strong statement since NERSC had long invested in Lustre (scratch) and GPFS (community). Very cool to see NERSC not settling for the status quo […]

Scott Atchley, who co-keynoted #isc25, posted a really meaningful response to my ISC25 recap blog post on LinkedIn (https://www.linkedin.com/posts/scottatchley_isc25-olcf-frontier-activity-7345786995765395457-lGoq). He specifically offered additional perspective on the 20 MW exascale milestone […]

Happy Canada Day, everyone 🇨🇦

Photos of a new, big, naked Cerebras cluster in Oklahoma appearing on the socials today. Pretty neat. Wonder if this is another G42 install.

This is an amazingly detailed yet accessible description of how tape storage (media, drives, and libraries) work. Even if you don't care about storage, the engineering that goes into making this all work is fascinating. https://dl.acm.org/doi/10.1145/3708997 (from […]

LLNL has an interesting vision of the future of HPC and workflows that aligns with a lot of what I heard at ISC: #hpc is no longer just the supercomputer, but the end-to-end services and ecosystem that enable discovery. The description is in Attachment (1) here: https://hpc.llnl.gov/fg-hpcc-rfi

Here's my notes from ISC'25 in Hamburg: https://blog.glennklockwood.com/2025/06/isc25-recap.html #hpc