The SC #136
Issue #136 of The SC. Weekly supercomputing news. Nvidia Vera continues to impress. Can agentic coding break the CUDA moat. Heterogenous compute gets real and AWS runs out of cloud.
I need a Vera CPU. Not because I have something I need to run that particularly demands one. No, not at all. Simply because the latest benchmark results only further confirm my suspicions that this is rather interesting CPU and I want to see how it stacks up for myself!
Tom’s hardware published a set of benchmarks for Vera last week. They have comparisons against Intel’s Granite Rapids but only 5th Gen (Turin) CPUs from AMD unfortunately. They also don’t include any comparisons to other ARM CPUs such as Graviton, Axion or Cobalt. Hey Jensen, hook me up with access to Vera and I can fix that for you 😆.
Slightly more seriously, and going on just the early data we have so far, I am very curious as to whether Vera becomes the first CPU where my answer to the question of “what CPU platform should I migrate my legacy 32bit x86 code to” isn’t something along the lines of “the extra effort for ARM probably isn’t worth it, just get it running on vendor independent x86_64”.
Using that as a nice segue into heterogenous compute, we had a couple of interesting stories last week further solidifying that theory. First up, Meta shared how they moved their AI training to AMD GPUs and used vibe (oops I mean agentic) coding to port the kernels to take advantage of the larger vRAM. I haven’t stressed it myself lately, but it does seem that we have turned a corner with AI assisted coding and if it can reliably help migrate code from CUDA to ROCm that certainly puts a chink in Nvidia’s armour. I really need to revisit our Intel x86 porting experiment I started earlier this year with Opus 4.5 and see how well it performs using current models.
Secondly, we saw Oracle announce that will be offering not only Google TPUs alongside Nvidia GPUs but also Cerberas and Positron. That heterogenous enough for ya? 😁
Apparently, AWS is sold out. Now I’m sure there’s a degree of spin behind that particular story, but I can also share that I’ve had a number of closed door conversations recently with clients, around ensuring they can continue to secure (often spot) capacity to run their workloads. This comes alongside a slow inching upwards of spot pricing (even on CPU compute) and a change in recommendation from cloud solution architects that pushes people towards multi region architectures. Things that make you go hmmm. That’s about as much as can be said here but if you need more data, you know where to find me. Bring an NDA.
We’re pulling together a new version of our financial risk analytics benchmark and I’d love your input. We’ve had some feedback already, but more is always better right?
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)
and Oracle starts providing Google TPUs as well as Nvidia GPUs
Is the cloud running out of capacity? This really isn’t just an AWS story either…
HMx Labs Updates
COREx needs an update. I need your input to get it right.
Plus Lucky has joined the team
Off Topic
My brain has been breaking lately around copyrights, trade secrets, patents and intellectual property in general.
I was supposed to write about this on Friday but that didn’t happen because I’m too deep into this rabbit hole right now so in the meantime you can jump in with me.
Of course, it started because of what’s happening right now but then it got messy as I started looking at this history and the research. :scream
Start easy with a YT video on the topic
Then dig in
- Can't block, must run: Small firms and appropriability – Byma & Leiponen
- Little patents and big secrets: managing intellectual property – Anton & Yao
- Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy – Teece
Or just cross your fingers and hope I put finger to keyboard on this topic this week and give you a TLDR along with my thoughts on where this goes next.
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
