By Hamza in Quantum Weekly — May 18, 2026

Quantum #118

Issue #118 of the weekly HPC newsletter by HMx Labs. Cerberas Systems IPO, Multipath reliable connections (MRC) networking for ever bigger supercomputers and more on K8 in HPC.

Whilst its immediate impact is going to be pretty limited for most of us, I guess the most important thing that happened in supercomputing last week was the Cerberas Systems IPO. At least as of Sunday evening, it seems to have been relatively well received. Whilst Cerberas may not have a particularly wide deployment yet and technical adoption is low; the financial markets at least seem to believe it has a future.

Which poses some interesting questions for the technical amongst us to ponder upon as it is yet another sign of increasingly heterogenous compute environments with very little underlying support by way of multi architecture frameworks. Most HPC developers, RSEs and the like will rarely target more than a single architecture, and this worked fine in a world where the supercomputer you were going to run on sat in your building and that was all the choice you had. As we move to a world where supercomputers are in the cloud and software defined, that approach may not work so well. Or maybe the public inference use case will move to the edge and what remains (trainings, classical MPI workloads, converged LLM plus classical compute) will be like a lot of HPC workloads and target a very specific architecture.

The second story to catch my eye from the last week was HPC Wire’s article on Slurm vs Kubernetes. I do wonder if my piece a couple of weeks ago had anything to do with that 😀. The article ends with a reference to Slinky, and whilst I think that Slinky (and its numerous counterparts in this space) are a good idea if you have to use Kubernetes (say it’s the only way to obtain compute capacity or the entire estate now runs K8) I really wouldn’t select it as an option unless I had that restriction.

Lastly, Next Platform’s piece on Multipath Reliable Connections for supercomputer networking is pretty cool! Definitely worth a read and I wonder if it might inspire similar solutions at a smaller scale too. The equivalent of what a Beowulf cluster is to a supercomputer if you will. Honestly, I’m even wondering what we could do with our tiny little HMx Labs test cluster.