The SC #126

There were a few interesting updates in the world of supercomputing last week, nothing life changing but certainly some things of note to keep an eye on. 

First up we get the new Linux 7.2 kernel release which is worth keeping eye on for supercomputing folk because it includes within its many updates the ability to use cache (we’re talking CPU caches here) aware scheduling. i.e. you have the ability to run code on cores where the local cache already has the data you need. For multi socket and NUMA systems (even on a single machine, but even more so at supercomputer scale) this gets pretty useful. It might be a little while before these changes trickle down to your favourite distro but at least you know they’re on the way giving you time to start updating your code to take advantage of it.

Microsoft Azure has finally seen sense and is allowing you to choose how you use the compute you rent. You can now enable and disable cores on a VM as you see fit as well as enable and disable SMT/ Hyperthreading. And not too soon! Whilst disabling SMT where it was already enabled was previously possible on Azure, it did require using a fairly poorly documented API which itself was only enabled once you raised a support request for it. I don’t understand the fascination CSPs seem to have with wanting to dictate how we use their machines. Price it how you like but let me do what I want with it. Come on AWS, let’s see the same change from you now. Let me enable SMT on your [CM]7a VMs.

Not new, but HPC Wire provided a fairly nice write up of scale up, vs out vs across networking (for AI) as practised at Meta. There’s been a bit of chatter around this over the last week, so it seems like a timely piece. To my FSI friends who are reading this and thinking, what’s new here we’ve been doing this for years (just not at quite the same scale) then understand, this why you need to start talking a bit more about what you do in the supercomputing space!

We also had a nitrous oxide fuelled, turbo charged version of Cerberus’ wafer scale inference engine drop last week. CXL made a bit of progress into something that’s edges from lab to reality and AMD dropped a bunch of pull requests to get its Ultra Accelerator Link open standard GPU networking protocol into the Linux kernel.

Lastly, if you’ve been missing the dulcet tones of Boof breaking down and explaining HPC concepts and haven’t yet found his new YouTube channel which is now unencumbered from the oversight of AWS give it a look.

And that’s yer lot. Have a good week all.

All the News in Depth

Cloud and vendor releases for Supercomputing & HPC (and AI too if you change the filters)

Linux 7.2 gets cut and includes cache aware scheduling. Tasty.

Microsoft provides better support in Azure for controlling SMT and active cores.

Nothing new, per se, here but it’s not a bad summary of scale up, out and across networking which seems to be getting a bit of attention lately. 

AMD pushes PRs to integrate UALink into Linux

Cerberas overclocks its AI inference

Want to use CXL? Maybe it's getting closer to reality with Marvel and Synopsys releases.

I know this is old, but I somehow forgot to include it my updates so far! If you managed to miss Boof’s new YouTube channel it’s time to hit Subscribe on it

HMx Labs Updates

Not much from us last week… I’m in summer mode so here’s a picture instead. Free HPC Club T shirt if you can guess where HPC NErD is or what's in the background.

Off Topic

Deepseek now provides a coding harness based on a plugin model.

More anecdata supporting AI generated anything not being all that awesome after all

Know someone else who might like to read this newsletter? Forward this on to them or even better, ask them to sign up here: https://cloudhpc.news