No one remembers history. When IBM and Microsoft partnered on OS/2, IBM got upset when a MS employee rewrote some code to be smaller and more efficient because “negative lines of code!”. Measure the wrong thing and you’ll get the wrong behavior. Incentives matter. (╯°□°)╯︵ ┻━┻

AI is turning scientists into publishing machines—and quietly funneling them into the same crowded corners of research. spectrum.ieee.org/ai-science-r...

Nvidia’s New #Rubin Architecture Thrives on Networking Some computations happen while data is enroute spectrum.ieee.org/nvidia-rubin... @spectrum.ieee.org

AI Coding Assistants Are Getting Worse Newer models are more prone to silent but deadly failure modes spectrum.ieee.org/ai-coding-de... #AI @spectrum.ieee.org

The problem with saying "just tier to HDD if flash prices are too high" is HDDs are scarce too. HDD manufacturers opted not to invest in expanding fab capacity b/c they'd never recover the capex given declining HDD sales. HDDs are a myopic investment right now. blocksandfiles.com/2026/01/09/s...

The Rubin platform uses codesign across six chips — the NVIDIA Vera CPU, NVIDIA Rubin GPU, NVIDIA NVLink™ 6 Switch, NVIDIA ConnectX®-9 SuperNIC, NVIDIA BlueField®-4 DPU and NVIDIA Spectrum™-6 Ethernet Switch — to slash training time and inference token costs. nvidianews.nvidia.com/news/rubin-p...

“AGI” has turned into a term of hype rather than a term with a precise meaning. Andrew Ng proposes a new version of the Turing Test, the Turing-#AGI Test. www.deeplearning.ai/the-batch/is...

NASA's SPHEREx is creating a 3D map of the sky every six months over a two-year period, delivering a sweeping view of the cosmos. Behind the scenes, teams from NASA JPL, Caltech, the NEID project, and TACC built the mission’s first fully automated science data pipeline. Learn more: bit.ly/4b7yufa

Nvidia spends $5B on Intel bailout, instantly gets $2.5B richer

📢 REX-IO 2026 Workshop: Call for Papers! 📢 I'm happy to announce the 6th Workshop on Re-envisioning Extreme-Scale I/O for Emerging Hybrid HPC Workloads at ACM HPDC 2026! Submissions are due March 31, 2026 (11:59PM AoE). 📜 🦖 Further info: sites.google.com/view/rexio/ #HPC #supercomputing

NVIDIA explains why they used FP8 as a scaling factor in NVFP and not MXFP-style E8M0. The idea is to preserve the largest value in each block. buff.ly/SjwVr6w (16:00 ff) We found this to be very true in a study and managed to recover some MXFP accuracy with micro-rotations: buff.ly/gXaCAhF

Schedule for #HPC, Big Data, and Data Science devroom at FOSDEM'26 (on Sun 1 Feb 2026, in Brussels) has been published: fosdem.org/2026/schedul...

A well written (mostly) non-technical #AI primer by @thedeadline.bsky.social www.hpcwire.com/2025/12/16/a...

AI’s Wrong Answers Are Bad. Its Wrong Reasoning Is Worse. spectrum.ieee.org/ai-reasoning...

NERSC recently did a wholesale replacement of its FDR InfiniBand storage fabric to RoCE. The IB was a greenfield installation back when I started in 2015, and replacing it with a competing technology in production is quite the feat. Glad to hear it succeeded. www.nersc.gov/news-and-eve...

IDC: $314.2B in GPU-accelerated systems in first 3 Q of 2025 How sustainable is this crazy server spending? ODMs are now pushing almost 60% of worldwide server revenues, up from 45% last year. www.nextplatform.com/2025/12/16/h... #AI #HPC @nextplatform.bsky.social

The best version of #HPC: NCSA director Bill Gropp on ‘imagining an HPC utopia’ www.eurekalert.org/news-release... @ncsaatillinois.bsky.social

NVIDIA has acquired SchedMD — the leading developer of #Slurm, an open-source workload management system for #HPC and #AI blogs.nvidia.com/blog/nvidia-...

#HPC in Transition: Jack Dongarra will deliver the #ISC26 Closing Keynote on Thursday, June 25, 2026.

The EuroHPC Federation Platform will bring @eurohpc-ju.bsky.social systems together, making it easier for researchers and industry to find, access and use computing resources across Europe from a single place. Next step: the launch of MyEuroHPC web interface in March 2026 🔗 csc.fi/en/blog/euro...

I wrote up my notes from #SC25. Have a look: blog.glennklockwood.com/2025/12/sc25... I’ll keep picking away at the editing, but would love to hear more from others about what stood out to them. I wasn’t at the conference itself as much this years as in the past, so I know I missed a lot. #HPC

A brave stake in the ground that defines what is (and isn’t) a parallel file system. I generally agree with Chris’ explanation. But I’m sure he’ll get hate from parallel storage elitists who don’t like how inclusive his take is.

DOE to build an integrated discovery platform by linking together #supercomputers and other facilities at its 17 National Laboratories with industry and academia drawing on the expertise of roughly 40,000 scientists, engineers, and technical staff. www.theregister.com/2025/11/25/t...

NVIDIA will introduce liquid-cooled busbars into the racks for its Vera Rubin platform, part of a broader evolution of data center racks and power design to support more powerful AI computing. open.substack.com/pub/datacent...

Intel: We’re simplifying the Diamond Rapids platform with a focus on 16 Channel processors and extending its benefits down the stack to support a range of unique customers and their use cases. (Source: Intel Spokesperson to STH) www.servethehome.com/intel-cancel... @servethehome.com.web.brid.gy

Silly thread for a Saturday: some of the #HPC clusters I’ve worked on over the years. First up is Cielo, a Cray XE6 I worked on at LANL! Which might actually be the prettiest supercomputer I’ve worked on.

A random interesting fact about the Cray XE6 racks: air exhausted up from the top of the rack! The result was that the lights above Cielo all started failing over time due to being blasted with hot air. And no good way to access them due to the size of the cluster. Clearly an early photo 😁

Exclusive: Amazon.com is joining Microsoft in supporting legislation that threatens to further limit Nvidia’s ability to export to China, a rare split between the chip designer and two of its biggest customers.

Exclusive: Samsung hikes memory chip prices by up to 60% as shortage worsens, sources say reut.rs/49QjvFI

Very nice overview of the emerging UALink standard with nice features such as splitting packets in switches, in-network computing, high energy efficiency, and lowest silicon overhead: buff.ly/AgLvC1g I'll be joining a panel at SC25 contrasting UALink and UEC next Wed: buff.ly/BeCMFcL Join us there

DARPA’s Next-Generation Microelectronics Manufacturing (NGMM) program is building a packaging plant in Austin that is dedicated to 3D heterogeneous integration (3DHI). spectrum.ieee.org/3d-heterogen... @spectrum.ieee.org @darpa.mil

If you can read an analog clock correctly, you are still outperforming #AI in that regard. spectrum.ieee.org/large-langua... @spectrum.ieee.org

Uhm, there is a typo in the headline, remove the "s" from insane. That should fix it.

Paderborn's new Otus #supercomputer features 142,656 processor cores, including AMD “Turin” and #Nvidia H100 #GPUs, and 5PB of storage managed with IBM Spectrum Scale (formerly GPFS) file system. ow.ly/mnlX50XqGAs

Andrew Ng: #AI has stark limitations, and despite rapid improvements, it will remain limited compared to humans for a long time. #AI is amazing, but it has unfortunately been hyped up to be even more amazing than it is. www.deeplearning.ai/the-batch/is...

Europe takes a major step in research connectivity! A new terabit network will link supercomputers across the continent, including EuroHPC’s @lumi-supercomputer.eu located in CSC’s data center in Kajaani 🚀 🔗 csc.fi/en/news/tera...

Are 2030 AI hyperscalars capital constrained, power constrained, DRAM constrained, flash constrained, compute constrained, software constrained, or :-) demand constrained?

Racks filled with GPUs and liquid cooling gear can now weigh 6,000 pounds or more, requiring new approaches to address human safety and investment protection. Google, Meta, and Microsoft are turning to robotics to safely move these huge racks. open.substack.com/pub/datacent...

Scammers have a new way of getting into your pockets: by targeting your #AI assistant. They use prompt engineering, embedding code in emails that trick AI tools into taking malicious actions. Learn how to protect your digital presence. spectrum.ieee.org/ai-agent-phi...

Can we build an #AI #Climate Scientist? Asked at the ADIA Lab Symposium in Abu Dhabi last week - now online at buff.ly/6igSeyg :-). Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.

AI excels in complex tasks but falters at reading analog clocks—what does this tell us about its limitations?

Nvidia's biggest scale up domain is 72 GPUs. Google's is 9,216 TPUs. Historically TPUs have trailed on FLOPS, memory, & bandwidth. That's no longer the case with Ironwood. Google has a Blackwell-class TPU with absurd scale. More on @theregister.com ⬇️ www.theregister.com/2025/11/06/g...

Exclusive: Intel is losing a data center AI executive who previously helped lead the company’s Gaudi accelerator chip efforts and is now headed for a job at AMD, CRN has learned. www.crn.com/news/compone...

Collaborator and friend Dan Alistarh talks at ETH about using the new NvFP4 and MXFP4 block formats for inference. Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks. arxiv.org/abs/2509.23202 Great collaboration and cool stuff

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 glennklockwood.com/garden/atten...

OpenAI spreads the imaginary wealth beyond Microsoft with $38B AWS deal Amazon deal still dwarfed by $250B Azure commitment made as part of OpenAI's for-profit transformation www.theregister.com/2025/11/03/o...

Silicon Valley’s biggest companies are already planning to pour $400 billion into artificial intelligence efforts this year. They all say it’s nowhere near enough.

The largest hyperscale operators say demand for AI services is filling data centers as fast as they can build them, with several saying they are compute-constrained. As a result, they expect to build even more data center space in 2026. datacenterrichness.substack.com/p/hyperscale...

Each time a new AI training benchmark is introduced, the fastest training time gets longer. Then, hardware improvements gradually bring the execution time down, only to get thwarted again by the next benchmark. Then the cycle repeats itself.

Diamond Blankets Will Keep Future Chips Cool Growing a micrometers-thick layer of diamond inside advanced chips spreads out the heat and drops the temperature more than 50°C. spectrum.ieee.org/diamond-ther... @spectrum.ieee.org