This article starts strong about how universities’ insulation from market forces gives them a good footing to explore long-term challenges in AI around things like explainability, but its three recommendations for NAIRR seem pretty weak.

technologyreview.com/2024/04/19/109…

My old colleagues at NERSC are hosting an end-to-end LLM bootcamp with NVIDIA which is open to the greater HPC community. Sounds really interesting. I miss being able to take time to participate in these sorts of training.

Runs May 7-9, details here: nersc.gov/users/training…

This article makes it sound like Microsoft was told to go buy G42, the UAE AI cloud, out so that G42 wouldn’t make an equivalent deal with China. Strange proxy war going on, with US tech giants in the middle.

nytimes.com/2024/04/16/tec…

I see we’re well on our way to making Terminators.

I’m surprised by how many people experiment with GPT on tasks involving shooting things with guns.

Glenn K. Lockwood I've read the articles by Sharada Yeluri for some time now as she writes excellent, well illustrated, explanations of networking.

Impressive to see a company that started with a strong dependence on exotic hardware (3D XPoint) pivot to being completely software-defined and cloud-ready in just a few years.

Incidentally, Azure is the last of the Big Three to get VAST platformed.

Things I didn't think I'd have to learn about this week in HPC/AI: VxLAN

Outside of scientific computing/physical simulation, what workloads rely on FP64?

Do any make productive use of quad precision?

Infrastructure-as-a-service circa 1960. Neat thread.

Dairsie Latimer Glenn K. Lockwood How many hours do you have?

(Beyond the meme, it’s mainly attributed to non monotonic rounding on A100 generation tensor cores for all types other than FP64.)

the whole thread, but esp this middle part

as a college student, I used to feel shamed by the “every EE/CS grad should know X” discourse, but now I know that most of those people have zero perspective on how vast our industry is and how short college is

Fun activity: Pick your favorite nuclear power station, find it on Google Maps, then look for nearby, massive warehouse-like building complexes with huge switchyards.

As AI booms, land near nuclear power plants becomes hot real estate theregister.com/2024/03/25/ai_… via @theregister

.HPCpodcast talks to Matt Sieger, project director for OLCF-6, the planned successor to the Frontier exascale #supercomputer

Objectives, status, timelines, along with some surprising possibilities regarding vendors

insidehpc.com/2024/03/hpcpod…

#HPC #AI via insideHPC.com

“We’ve been building up this #AI #supercomputing infrastructure capability for many, many years”

John Lee, Microsoft Azure’s Principal Lead, #AI Platforms & Infrastructure, predicts hardware bifurcation between training & inference

insidehpc.com/2024/03/advanc…

#GTC24 #HPC via insideHPC.com

It will be my great pleasure to present this award at ISC HPC. This is a great paper on a great topic.

This article is a few months old, but still offers rare insight into Microsoft’s big AI supercomputers. news.microsoft.com/source/feature…

Nvidia Blackwell rack - 130 kW/h of power and exaflop of fp4 and one node with water cooling 😻 #GTC24 #HPC is totally swayed by AI now… who cares about fp64 and fp32 😿

From the LinkedIn post: “Look what I saw parked out in front at GTC…”

#yikes