The SC #121
Issue #121 of The SC. Weekly supercomputing news. TSMC invests in more capacity, Nvidia claim they are on track with NVL144 and is quantum computing starting to cement its position as the next accelerator?
There seems to be concern that Nvidia is going to slip on its delivery of its Kyber NVL144 racks, even as Jensen continues to promise that the roadmap is intact and on track. There’s a couple of interesting things here. Specifically, it’s not the production of Vera CPUs or Rubin GPUs that’s in question here, it’s the production of the NVL144 rack and the additional interconnects this entails.
Firstly, this gives competitors like AMD and their Helios rack or Google’s TPU architecture a bit of leeway to make inroads into Nvidia’s market. Critically, what’s going to be important here is not the performance of a single GPU but the scale out capability of the entire rack.

Which brings me to my second point.
The first supercomputers, such as the Cray One used its own integrated circuits designed specifically for it, the evolution and market dynamics of integrated circuits (of which you CPU is one) meant that later supercomputers, such as El-Capitan, were built on integrated circuits provided by companies dedicated to building them such as AMD and Intel. It was companies such as Dell, or HPE that then took those CPUs coupled them with cooling, networking, racks etc and created a supercomputer. Cray disappeared and became part of HPE.
Nvidia is now creating not only the silicon, but the entire rack. I’m not sure if that’s the silicon manufacturer eating the ODM but it does mean we’re returning to an era where the same company is creating the silicon, the interconnects, the rack design and even part of the cooling infrastructure. I wonder how this plays out.
Talking about interconnects, Marvell has released its Teralynx 100 kit.
TSMC is taking its massive profits and investing not only in additional capacity but also the USA. I guess if you’re biggest customers are spending like crazy now is the time to expand.
And coming back to what the ODM now do, they seem to be turning to integrating other accelerator options, specifically quantum, into their offerings with both HPE and Dell offering something in this space and placing quantum computing firmly in the realm of your next accelerator as opposed to your next data center.
All the News in Depth
Cloud and vendor releases for Supercomputing & HPC (and AI too if you change the filters)
TSMC invest further in the USA and plans for greater expansion, or see Next Platform more their more in depth take.
Where’s that Kyber NVL144 gone?
Marvell ups its game in supercomputing interconnects
Quantum is firmly establishing itself as the next type of accelerator
HMx Labs Updates
HPC is like Poker. You always need to know who’s bluffing

Compute futures and markets I suspect would be amazing for HPC and supercomputing. We have them. Sort of. Sadly not in any way that’s actually useful to the people that need to consume the compute

Off Topic
Open source is getting harder and harder
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
