The SC #130
Issue #130 of The SC. Weekly supercomputing news. A cool idea to solve the memory wall. Nvidia gives us better power management. High speed networking (UAL) takes a step forward. HPC Club date set.
The most interesting story last week wasn’t details on a new CPU or accelerator, a new product version or cool new startup with a solution in the supercomputing space. It was just an idea, a thought experiment. Anyone working in HPC for more than a hot minute (so non of you AI-come-lately types then 😆 … I’m kidding, relax already!) knows that raw compute hasn’t been the problem for quite some time. The memory wall is all too real. Jack Dongarra even talked about it in his closing remarks at ISC 26. If modelling memory (even if that’s only in the L3 cache) as something other than a linear block can solve that, I’m game to try it out.
The DOE (USA) is in a real rush to get some real-world science use out of quantum computing and has put up a cool $215m (what’s that, half an NVL144 rack these days?) in a competition. Shame we can’t get some of that if we manage to get financial risk analytics working on quantum huh. 🤣
Nvidia had a couple of interesting announcements (and neither of them are about GPUs). First up they presented DSX, which, strictly speaking, is a little boring because it’s just about getting various data centre utilities to talk to each other so they can all operate at the right level and use less power. Overall, that’s kinda cool though, right? I have to say I’m surprised Nvidia didn’t just try and solve that by throwing a swarm of agents at it and getting them to just control all the knobs in the DC. I am quite glad they didn’t though!
The second release is also power related, but this time around integration with the grid to redirect compute to where power is available. Again, also kinda cool. If/ when it works it will be at least anyway.
We also had Cornelis and Delos Data both release products to provide an alternative to Nvidia’s now closed networking based on its Mellanox acquisition. I’ll admit, I do have my fingers crossed some of this stuff works out.
A little less notice than we normally like to provide but we have finally set a date for the last HPC Club for this year. October 14th and it will take place in London. We should have the booking link live on Wednesday morning so keep an eye out for that if you’d like to come along, If you want one of the fancy new T shirts then you’ll definitely want to come along (and maybe try and bribe Mugheesa to make sure she orders one in the right size for you).
Have a good week all.
All the News in Depth
DOE lines up some serious cash to accelerate quantum computing adoption in real world use
Nvidia gets DC infra to communicate to help save power
Couple this with a standard for getting DCs to talk to each other and the grid…
Supercomputer interconnects other than the old Mellanox gear that Nvidia now owns start picking up
HMx Labs Updates
The next HPC Club will take place on 14 Oct in London.


Off Topic
Scaling LLMs. I guess we all kind of knew this instinctively but now MIT has maths that says it too.
Just in case you’ve been living under a rock and haven’t seen the original
The modern life of a software developer getting you down? You’re not alone
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
