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

You’d be forgiven for thinking that if you exclude all the mainstream AI news, especially everything that’s more about financing and power grids that nothing much of note happened last week. Even my AI research bots turned up little to write about. A little old fashioned manual digging in release notes did turn up a couple of interesting stories though.

We have word from Microsoft that that they will not only be releasing an update to their own AI chips (Maia) but also rolling out a surprisingly aggressive number of them. Treat that statement with a pinch of salt but it does rather confirm the very heterogenous nature of compute. Yes, present tense. Now.

Nvidia gave us an open source LLM router which I suppose could be interesting in some domains of HPC to try and control cost where we have converged LLM inference and classical compute. I’m not sure what the workflow for that looks like though. I guess you need the routing decision before the scheduling decision else you can’t allocate resources correctly. I wonder how long it is before we see an LLM router integrated into a HPC scheduler. For LLM inference only workloads this seems like a bit of a no brainer surely? 

How is it possible that even with AI genies writing code at a rate that eclipses my fat fingers on a keyboard that I still have more possible ideas than engineering capacity. 

Riddle me that batman. 

Talking of converged AI and HPC compute, Google gives us AlphaFold for HPC as an addition to their catalogue of offerings. Looks like an out of the box way to run AlphaFold on Google Batch to reduce your compute wall time.

AWS also gave us Kubernetes style health checks for the application layer in EC2 (see the release notes in Noteworthy). Please let this be another nail in the coffin of HPC on K8.

Lastly, we welcomed Eduardo to the team last week and I had a bit of a rant about synthetic and natural intelligence.

All the News in Depth

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

Microsoft is still investing in its own Nvidia alternatives for AI focused compute with the next generation of Maia processor due for release imminently.

Nvidia provides an open source LLM router

Google adds to their converged AI/HPC capability with AlphaEvolve for HPC

HMx Labs Updates

Not sure if this is a rant, man shouting at clouds or just incoherent mutterings of a mad man. I can say none of it was written, edited or illustrated by AI. I think I had a few interesting things to say but do feel free to disagree

AI: Context, Taste & Inspiration

Somewhere between a rant and some random thoughts on AI.

Flux by HMx LabsHamza

Also, welcome to Eduardo!

Eduardo Joins HMx Labs

Welcome to Eduardo Visinheski who joins us at HMx Labs as our newest HPC Engineer.

Flux by HMx LabsHamza

Off Topic

AI progress is accelerating on an exponential curve right? Right? Maybe not?

Is the Dunning Kruger effect just nonsense? I don’t know, but I can say I agree with Sabine’s last point, that data cloud doesn’t look like it has any kind of correlation at 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

Time to go off topic a little and stray into a mass market subject. AI, context, taste and inspiration. I’m not sure I have anything coherent to say but I had a few thoughts knocking around in my head that I thought I’d share. Consider yourself forewarned.

First up, context. Frontier models (hate that term) have a context window of 1 million tokens. That sounds pretty impressive. It’s not. Ok I guess compared to what was available a year ago, it is. Compared to the hyped claims of AGI its pathetic. Plenty of people have already explained that it’s too small a context window to hold even a moderately sized real codebase. I don’t want to talk about that.

A decent code harness can mostly work around that limitation by grepping through your files and reading the bits it thinks match. Honestly, that’s often good enough and if you don’t know the code its certainly faster than attempting to do something similar manually. 

Know what it can’t do? Hold 20+ years of experience in its head. Or even the 2+ years of a junior developer. It doesn’t know your company’s infrastructure restrictions or regulatory requirements. Things even your most junior developer picked up in the first week and stuck with.

This often gets banded about as “taste”. Another stupid word that’s being used to hide the fact that it’s just more context. Context that the model doesn’t have access to. Context that the model can never hope to hold in its context window. Taste is a stupid word. It’s just knowledge. But if we call it that, it belittles the AGI narrative. 

Taste is just context your model can’t access. 

Want to stay in a job? Make sure you always have more context than the AI.

Still think that AI can be creative? Only if your idea of creativity is reversion to the mean. Let me try and illustrate this with a bit of anecdata and I won’t use code for these. Earlier this week I wrote a post introducing our newest HPC engineer. When I sat down to write that post I had the first three lines only. This was all I could come up with:

Just a quick note to welcome our latest HPC engineer, Eduardo to the team. Welcome Eduardo! That’s all folks!

I sat there with just that on the screen in front of me for the longest time. Then I cracked. I ask ChatGPT 5.6 High. Yea. I wasn’t about to mess around. Throw the most capable expensive model at it right away. I told it to write the post in my own style, and this was what it came up with:

Very pleased to welcome Eduardo Cook Visinheski to HMx Labs as our newest HPC Engineer.

Eduardo joins us to help work on the increasingly difficult problems that sit somewhere between HPC, cloud infrastructure and “why is this behaving differently on that machine?”

Exactly the sort of problems we like.

Welcome to the team, Eduardo. Looking forward to building some interesting things together.

🤮

If I ever produce anything quite so bland please take me out back and give me a good kicking till I see sense. I thoroughly admonished the poor 4 bit matrix math addled simpleton and off it went toread more of my previous ruminations and managed to spit out the following:

“HMx Labs has a new HPC Engineer.

Welcome Eduardo Cook Visinheski 👋

We’re building a team around a fairly simple premise: there are a lot of interesting problems in HPC that still aren’t solved particularly well.

Cloud. Scheduling. Performance. Portability. Heterogeneous compute. And computers occasionally producing different answers because you looked at them funny.

Eduardo has decided to join us in trying to fix some of them.

Excellent decision. Possibly.

Welcome to HMx!” 

Better. Still not me. And critically, still garbage. I’d never put my name to that.

Here’s where I will give it some credit. After having been stuck on a largely blank page for an hour at this point, the above did manage to trigger me to write the following paragraph.

“Eh. You’re still here? Ok fine. Eduardo’s made the mistake of letting himself be duped into solving the kinds of problems in supercomputing that everyone else would rather pretend don’t exist. The ability not just to just bet on the price of compute but to actually settle the trade with delivery. Figuring out what a next generation scheduler that can deal with HTC levels of throughput but blend it with topological awareness for AI inference as part of the workload. Transforming switching schedulers from a multi-million pound multi-year change programme to a deployment and the flick of a switch. You know, simple stuff 😆”

I think it’s better. Perhaps you disagree.

Let me give you another example. I like taking photos. I seemed to have not taken many lately and the last real camera I bought was in 2012.  A rather large professional grade DSLR and I had been happily using it since then. The mirrorless revolution had largely passed me by. Using ChatGPT and/or Claude to quickly research cameras, aggregate real user feedback (and not just shill reviews) was actually pretty great. Though I still ended up reading far too many reddit threads. You know what sucked though? The little bots ability to actually recommend something.

Despite having a huge amount of context in a back and forth that spanned months, without fail the recommendations were always close to what most “The Best Camera” web searches would produce. Which in turn was mostly human and AI generated sponsored slop lists with little to no substance.

I wanted a “proper” camera, with manual controls, a decent sized sensor and enjoyable to use. But it had to be cheap enough to not worry about throwing into bags and taking down mountain bike trails, snowboarding and using on the beach. 

The AI recommendation? Olympus Tough TG-4. Fair enough. Except I’d hate using that. It doesn’t meet any of my requirements. Even with further prompting and having read, by this point, literally thousands of Reddit threads exploring cameras that align to I like in a camera, it was incapable of recommending something that actually fit what I wanted. 

Because what I wanted was out of the ordinary. What I wanted was bit left field. An old, but therefore cheap enough to not care, Sony NEX5. Or a discontinued series of camera’s by Nikon, the 1. The AW1 more specifically as it is waterproof and impact resistant. Not much chance of AI coming up with either though.

This pattern repeated with multiple use cases, so it wasn’t a one off either.

Don’t trust the genie to be creative for you. But it might spark an idea in your own generative intelligence neurones.

Which brings me to my last point. I don’t care if you’re coding, writing LinkedIn posts or composing symphonies. You need to train that creative muscle. Heck even actual body builders understand this concept. It’s called “greasing the groove”. The act of repeatedly exercising certain parts of your brain leads to strengthened neural pathways. It doesn’t matter if you’re doing pull ups or writing HPC code. If you’re not actively exercising your brain to do it, someone else who is, will take your job. 😁

Just a quick note to welcome our latest HPC engineer, Eduardo to the team.

Welcome Eduardo!

That’s all folks! 

Eh. You’re still here? Ok fine. Eduardo’s made the mistake of letting himself be duped into solving the kinds of problems in supercomputing that everyone else would rather pretend don’t exist. The ability not just to just bet on the price of compute but to actually settle the trade with delivery. Figuring out what a next generation scheduler that can deal with HTC levels of throughput but blend it with topological awareness for AI inference as part of the workload. Transforming switching schedulers from a multi-million pound multi-year change programme to a deployment and the flick of a switch. You know, simple stuff 😆

Hopefully he doesn’t run away now that he knows what he’s in for.

And no, the photo is NOT AI. I took it! I’m looking at you Omar 😡

Quite an interesting week for supercomputing news, albeit a little AMD dominant though of course it’s hard to talk about anything GPU these days without Nvidia being mentioned in the same breath.

Remember Talaas? They had a fair bit of fanfare earlier this year with AI inference token throughput rates that were off the charts. AMD just bought them. So now Nvidia has Groq, AMD has Talaas. As I’ve said in the past, when a use case becomes sufficiently prevalent it ends up being baked into the silicon. LLM inference will be no different and that’s across consumer, edge and data centre. Or at least I think so.

When we released our scheduler selection tool last year, I didn’t think we’d need to make many changes. I mean, the HPC scheduler market seemed pretty mature, and companies weren’t exactly falling over themselves to create new ones now were they. Yea how wrong was I. AMD just dropped one too. ROCm Spur. Obviously, its targeting GPU scheduling as the primary use case, is Slurm compatible and of course written in Rust to keep up with the cool kids. Having said all that, I’m not sure what meaningful addition this brings to the table. Guess we need to update the scheduler selector tool again!

Want to do something with your GPUs that isn’t just 4 bit cat picture generation and needs real precision. AMD has always been supportive of this use case with Nvidia pushing emulation using the Ozaki scheme instead and still hoping to win the top trumps game. Looks like AMD isn’t messing about. Their native hardware FP64 support in MI430X GPUs weighs in faster than Nvidia’s Rubin GPUs even including Ozaki emulation. Crikey. I do wonder how much longer that CUDA moat will hold out.

The scientific community seems to agree with the new DOE supercomputers being AMD based.

Let’s talk CPUs for a minute. Nividia’s new Vera chips seem pretty great, just not quite as great as their marketing material would have you believe. I’d quite like to get my hands on one to see how it stacks up for financial risk analytics workloads rather than the nonsense of Agentic AI that people seem to think needs an insane amount of compute.

All the News in Depth

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

Vera is a pretty good CPU; shame they are marketing it as though it isn’t and resorting to fairly petty things.

AMD buys Talaas the AI inference ASIC company. Reuters give you the quick take but if you want meandering tale the ends with surprisingly little detail about Talaas but gives you lots of context then the Next Platform article might be more your speed.

AMD is also rattling Nvidia for real FP64 performance. It was never a question if you needed real non emulated FP64 but even if you can accept Ozaki emulated FP64 it looks like AMD still comes out ahead!

AMD has even dropped their own, Slurm compatible, scheduler. Honestly I’m not sure who this is targeted at or how it moves the game forward if at all.

HMx Labs Updates

We’re starting to look at what version 2 of the COREx benchmark should look like. Want to provide some input?

Benchmarks Should Represent Your Workload: COREX v2

This is your reminder that a benchmark is pointless if it does not represent your own workload. Period.

Flux by HMx LabsHamza

We’d also like to know which venue you prefer for HPC Club

HPC Club: Which Venue Did You Prefer?

What’s more important to you, a casual setting for the conversations or a better environment for the presentation?

Flux by HMx LabsHamza

Off Topic

Most open source projects you’ve heard of probably aren’t developed by lone developers tolling away under the midnight oil out of the goodness of their hear (though for sure that exists). No most of them are probably paid handsomely by big tech. Ever wondered why? This is a pretty solid take which I mostly agree with.

This was quite a good watch.. and I certainly agree that if you take away the noise then there really has been very little of note in the AI space lately.

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

Some of the big chip manufacturers are making lots of noise about how great their silicon is on various benchmarks and attempting to bludgeon you into belief with chart crime and long words. Fortunately, we have some actual brainiacs calling them out on it too (not me).

So, this is just your friendly reminder that benchmarks are meaningless unless they genuinely reflect your own workload. It really is as simple as that.

Which brings me to COREx, our little financial risk analytics benchmark. The same absolutely applies to this too. Which is why we’re working on version 2. 

Financial risk analytics is not a homogenous blob. When I first created COREx, the next best thing was running the unit tests for QuantLib and this was certainly an improvement. 

But we can do better. 

Specifically, I feel the difference between real time risk and end of day risk needs to be better captured as does the difference between market risk and portfolio risk.

That’s just me though, so before we run headlong into this, I’d love to get some input from you too. 

I have a question for the HPC Club regulars amongst you. 

Especially if you came to the last event at The Crypt and have been to previous events at Notes Café. Simple question. Which venue did you prefer?

We broke with tradition a little for the last event to allow us to have a remote presenter (I hope you’ll agree it was worth it), but if we stick to our general rule of in person only for the next event, where would you prefer we hold it?

Personally, I think The Crypt was amazing for the presentation. Having decent audio, a big screen and seating setup for a presentation made a huge difference. But I also like the really casual nature and easy conversations that happen when we hold it at Notes.

Which did you prefer?

AI infrastructure made lots of noise, mostly I guess because its earnings season, IBM and partners claimed quantum advantage and the EU wants to join the AI party.

Let’s talk quantum advantage first though because I’m sure you had enough other outlets pushing AI news down your throat last week. IBM and partners, announced quantum advantage for several use cases with IBM’s 156 qubit Heron. The claims here are fairly impressive with validation of the quantum compute using other trapped ion systems from Quantinuum and the classical compute comparison done using Fugaku which held the number one spot on the Top500 in 2020 and is still ranked at number 9.

HPC Wire seem to concur that we’re close to utility, if not there already, for quantum computing too.

I’d love to provide some kind of independent commentary on the above, but this is really hard to do. Unlike, say AI, where it’s pretty easy to validate its capabilities for yourself. the same cannot be said for quantum computing. Even if you had access to a real quantum computer. The amount of work, and level of both domain knowledge and math required simply places it out of reach for most people. I would like to answer this question though for a use case that we work with extensively. Financial risk analytics. So, we’re working on that, stay tuned for more.

Its earnings season on Wall Street and of course because the biggest companies work in AI infrastructure which is the most super of supercomputing you’re probably impacted if you have to purchase any kind of compute let alone a new supercomputer. 

My condolences if you’re in that position. 

The news is not easy to decipher. On the one hand, we have worsening credit spreads on corporate bonds for AI infrastructure companies and Oracle being downgraded to just above junk bond status. On the other hand, we have stellar earnings and increased investment into AI infrastructure from the hyperscalers.

In the midst of this, the EU has decided to enter the fray with EuroHPC putting out RFPs for AI gigafactories, or in normal parlance, new supercomputers.

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)

IBM and partners claim quantum advantage

HPC Wire also seems to think real use of quantum may be closer than we think

Intel is increasingly looking like its back from its near death experience even if its CPUs don’t quite hold up to its x86 competitors

CoreWeave financing looks increasingly shaky

Oracle downgraded to BBB rating

More circular financing? Nvidia backs OpenAI’s to the tune of $250B

And the EU wans in on the action too

HMx Labs Updates

Is finding developers who understand memory management getting too hard? I’m not just asking as a theoretical conversation starter but because we need people who can and there are still lots of systems in financial services that need these skills

HPC Software Engineer Wanted

We’re still looking for a software engineer who can write code that manages its own memory… Know anyone?

Flux by HMx LabsHamza

Want to accelerate your HPC network (people you know rather than packets on the wire)? For less than the price of a sushi dinner? Read this

Want to Learn HPC and Build an Amazing Network?

The HPC Leadersip Institute, 2026 Edition is taking place at the Barcelona Supercomputing Centre again this year in September.

Flux by HMx LabsHamza

Off Topic

I like Alberta, she’s funny. Makes a valid point here too, that I’ve also made in the past is a much less funny way

The whole AI is a scary cyber hacker story from the last week or so is getting out of hand, here’s a good sane take on it all from Cal Newport. Search for it on your podcast app of choice or put up with the ads on YouTube. “The Real Story Behind OpenAI’s Rogue Model”

Codeberg approves the change to ban AI Slop code. I’m wondering if we need to move our private repos off GitHub at least and wondering about Codeberg or internally hosting Forgejo.

Lastly, the family and I are hoping to get away to the alps again for a spot of downhill biking so I’ve been watching the incredibly talented Danny MacAskill with them.

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

Psst. Over here. Come closer. I have something to show you.

At the risk of sounding like a scammer, what if I told you that there’s a way that you can be in a room filled with centre directors of national labs, heads of supercomputing and HPC people so smart it will make your head spin.

What would you give to be in that room? For three days. To be able to pick their brains and learn from their experiences.

You’re probably thinking of numbers with four or five zeros at the end? It probably should cost that. But it doesn’t. It will cost less than good sushi dinner. Yea seriously.

To cap it off you also get to visit the most beautiful supercomputer in the world. And its not just me saying that. It’s official.

By now you’re wondering what kind of scam this is and expecting me to give you my wallet address to send me some bitcoin right. But you’d be wrong.

Just go sign up on the Barcelona Supercomputing Centre website for the HPC Leadership institute. Officially they aim to teach you about things like procurement, TCO modelling, investment cases. And they do. But you’ll come away with much more. Including a network of peers that most people would literally pay thousands for.

I attended last year and am tempted to go again, honestly, it’s worth it just to visit Mare Nostrum and be in Barcelona with fellow HPC NErDs for a few days. 

7^th to 10^th September.

HPC Leadership Institute - Edition 2026

HPC Leadership Institute - Edition 2026

BSC-CNS

Is it just me, or is it actually getting hard to hire engineers how to write code that has to manually manage memory?

We’re still hiring for this role. 

I’ve had the odd conversation with some our clients too who seem to be struggling similarly in hiring for roles that demand memory management.

Curiously even developers who have worked with C++ for years, but only ever used smart pointers, seem to struggle.

Maybe this will be a dying breed of programmer, alongside COBOL developers.

Anyhow, we’re still hiring and if you know someone who knows how to write code that expect to manage its own memory, point them my way please!

LinkedIn Job Post

HMx Labs hiring HPC Software Engineer in London Area, United Kingdom | LinkedIn

Posted 9:23:04 AM. Despite the big talk from the like Messrs Altman and Amodei and product hype from Codex and Cursor…See this and similar jobs on LinkedIn.

LinkedIn

Finding the real supercomputing news this week and picking it out of the AI haystack was harder than usual. Which is odd because AMD released its Zen6 CPUs and MI4000 series GPUs.

In a week filled with news about increased spending from OpenAI and Google, the OpenAI and Hugging Face saga and declining OpenAI revenues, what would normally have been a fairly exciting release seems to have been almost lost.

The press release from AMD focuses on its Helios rack scale product but honestly, for those of us that aren’t hyperscalers or neoclouds (in other words most of you reading this) the more interesting part is the chips inside and putting them in your own standard 19” racks not rebuild your entire DC. Unless of course you’re about to replace your supercomputer at a national lab 😉

This is especially true because AMD are still producing GPUs with native FP64 that hasn’t been crippled. You know, stuff for doing real work 😁

I know what you’re thinking though, CUDA. You can’t use AMD. No idea how well this stuff works yet, but they also launched ROCm.ai to try and bridge that gap.

Nvidia being Nvidia though, couldn’t let AMD have the news cycle to themselves and released a few more details about Vera Rubin. We’ve been talking about that for so long, can’t we just get on with it and release it widely please! Maybe no one has any space left to install them (if Ed Zitron is correct).

Not exactly new, but Next Platform has an interesting piece this week too about using AI for quantum error correction. I’d say something snarky about this, but I will hold fire till we complete our literature review of using quantum computing in financial services (specifically for risk analytics). If that goes well, we may have something more interesting to talk about and if it doesn’t then I guess I have more ammunition for my snark cannon. Or more work to do!

In other news, we had our first full week at HMx Towers with Sahana having joined us as our new Strategy and Partnerships lead. We still have an open engineering role though, know anyone that might be interested?

All the News in Depth

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

AMD gives us new Zen6 CPUs, MI4000 series GPUs, including some that can do real FP64 work!

And HPC Wire’s take on the release

What about that CUDA moat? Yea AMD’s been working on that… 

Nvidia wasn’t going to let AMD take all the news attention though and provided a few more insights into Vera Rubin

Using AI for quantum computing error correction

HMx Labs Updates

We finally hired someone for our open strategy and partnerships role!

Sahana Christie Joins HMx Labs

We’re excited to welcome Sahana to the HMx Labs team! She’s new to HMx and also to HPC but brings with her a wealth of experience in strategy, comms and business development.

Flux by HMx LabsHamza

We still have one open engineering role though

Hiring: HPC Engineer

We’re looking for another HPC software engineer! Know someone that might be interested? Fancy pointing them my way please? This video was brought to you by the letters H, M and X and the number 42. Have a good weekend everyone. 0:00 /0:03 1× Apply on LinkedIn

Flux by HMx LabsHamza

Off Topic

AI spending seems to be accelerating still...

Even Google wants to spend more since it sold more cloud

Even though AI still can't make any money, even with ads

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

HPC is about the people. So, I’d like to introduce someone new, Sahana Christie, who has joined us recently. Not only is she new to HMx Labs but also to HPC.

She’s had a bit of a whirlwind start with us. I onboarded (and met for the first time in person!) her on a Monday afternoon at Heathrow airport as we both waited to board our flight to Hamburg for ISC 26. A baptism of fire perhaps. One she needed a well deserved holiday to recover from and she jetted straight off to Japan once we arrived back at Heathrow later that week. As I caught a taxi home, she casually walked across the terminal and boarded another flight 😄

Sahana has a strong background in strategy, business development and communications, words that I barely understand let alone know much about so we’re incredibly excited to have her and help us in this space.

We’ve made a deal where I’ll teach her about NUMA placement of workloads and L3 cache optimisation and she will explain to me what people mean when they say things like “align synergies to create additional stakeholder value” and “collaborate on go to market strategies to capture additional opportunities”. I think she’s got the short end of the stick 🤣

I guess sometimes HPC means high performance career.

Apologies for the picture, I didn’t one of the real Sahana so you get a lego version instead for now.

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

Not quite the point, but this is good example of hybrid AI/Classical workflows establishing themselves

HMx Labs Updates

HPC is like Poker. You always need to know who’s bluffing

HPC is like Poker

My funniest conversation from HPC Club: HPC is like poker!

Flux by HMx LabsHamza

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

Compute Futures are Here. Sort Of.

We now have compute futures, something I’ve been wanting for year. But they suck.

Flux by HMx LabsHamza

Off Topic

Layered silicon!

AI Rewrites can work?

Open source is getting harder and harder

Want some cheap GPUs? Almost.

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

I have been talking about compute futures for a long time. Since before it was trendy. Before AI drove compute demand through the roof. World famous financial exchanges and back street prediction markets alike are now offering compute future. And I am… underwhelmed.

Why? Surely, I’ve been wanting a way to trade compute and now I finally have it, right? Wrong.

What we have today is the ability to speculate about the price of compute. I understand this is one side of what financial markets offer. It’s not the side that I care about. What I am, and always have been, looking for is the side of financial markets that are used by people that actually care about using the underlying commodity. Not just betting on derivative.

Details are rather scant on Kalshi’s new compute futures and given that half the information I’ve read about it doesn’t even understand the difference between a futures contract and a forward I’m not going to hold my breath that this becomes something useful for anything other than speculation.

If what you really care about is actually buying and consuming compute and want financial markets for all the benefits they bring to that, we still have several unsolved problems that everyone is trying to sweep under the rug. Who care about physical delivery or settlement if all you want to do is provide a cool new replacement to meme coins.

I don’t need to provide the world with another thing to bet on. I want to provide something that has real utility for those us that need to use that compute.

This means solving problems like delivery. No one that is serious about consuming large amounts of compute today is buying a futures contract on a prediction market because their use of that compute will be tied to some very specific APIs and network connectivity. We have nothing even approaching a standardised way to consume compute let alone arranging actual delivery.

It also means solving the fungibility problem. The performance of the same cloud VM type across regions can vary by over 20%, the difference in performance across vendors is even greater. A compute future for a H100 is a completely nonsensical product if you actually care about using that compute and not just speculating on the price of it. And no, the solution here isn’t to make everything homogenous. It’s to know how to have the right exchange rate. An exchange rate that varies by workload.

I’m not just ranting about this either. These are problems that not that hard to solve. We are doing our little bit (see ORB and FLOPx). Doesn’t make it any less frustrating.

HPC is like poker.

I don’t recall how exactly this conversation came about, but that was the opening claim. As per the rules of HPC Club, I won’t disclose who said it, but that person is welcome to claim credit here 😄

“HPC is like poker. You start of thinking it’s about the cards but then realise it’s about the people.”

The idea being, of course, that with HPC it’s not really about the hardware specs, how many GPUs you have or the scheduler you run. It’s about understanding and enabling the users to run the workloads they need to. That wasn’t my response though.

“What, you mean  you need to able to figure out who’s telling the truth and who’s bluffing when they ask you for 256 GPUs?”

I think this bit just as true!

P.S: HPC Club is great to pick up free swag too 😁 🥳