What are your go to benchmarks when selecting hardware? Both for GPU and CPU workloads?

We have a few changes we’d still like to make to FLOPx, our benchmark data platform, but I’d like to start extending it to include other benchmarks too. To have data that goes beyond what we have used already to serve our existing clients (which is very financial services focused).

This probably means open source benchmarks but I’m open to including proprietary/ closed source benchmarks provided they can be licensed and meet our criteria on transparency and repeatability.

What else would you like to see? The QuantLib test suite? Something based on OpenFOAM (is there a standardised repeatable benchmark already out there rather than just running one of the samples?).

I’ve been largely thinking of HPC related use cases, but I guess this is more widely applicable? Should we include some basics like CoreMark?

Friday afternoon musings on a post card please 😄

I first heard the idea of using LLM inference as paths in a Monte Carlo simulation at the QuantMinds conference (November 2025). The same idea has now been suggested by people from the leading AI labs including Andrej Karparthy. Just in not so many words.

Last month Andrej Karpathy release Auto Research. The idea being promoted is that an LLM, given tight validation criteria, can take a task and iteratively improve on the results and throw away anything that is wrong or produces a worse result. Not long ago I was at a private event with Anthropic presenting on the best ways to use Claude Code. One of the central points (of four) here was the importance of automated validation of the output of Claude Code.

So far nothing new here. I think those of use that have been using LLMs regularly independently came to the same conclusion too. So, we’re all on the same page, right? Almost. There’s a subtlety I missed here that I think the quants saw a year ago. I was just dismissive of it because of the cost of inference.

What no one is saying, is that if you have to gate the output of your “AI” on a validation step, is your “AI” actually intelligent? Or do you just have a better random number generator for your Monte Carlo model?

Let me try illustrating the point, probably poorly, with an example. It is possible to calculate the value of Pi just by defining a circle and generating random numbers. Draw a circle on a square piece of paper and throw 10,000 darts at it. You can now calculate the ratios of the areas (and therefore the value of Pi) by counting the number of darts inside and outside the circle. 

Imagine now if instead of a single circle we had two, concentric circles, one slightly smaller than the other. Like a car tyre. We can adjust the math to calculate Pi based on the number of darts that hit the tyre. The result is actually worse (less accurate) than the first method. But what if now we change the random number generator so it’s not random. So, it throws the darts mostly around the tyre. The accuracy improves (for the same number of darts or Monte Carlo paths). I have some half-baked ideas of how this might apply to equity option pricing but I’m not a quant so they’re probably wrong. Give me a shout if you are a quant though and want to talk.

This analogy though is what we have with LLMs. A better random number generator. That’s not really random. It might even be mostly correct. But not always. So you need less throws of the dart (or MC paths).

I think the finance quants just saw this way before I did. Before most of us did. Before the AI Labs will say it out loud but will quietly point to it. I thought it was mad because the cost of one LLM inference is many thousands of times that of running a single MC path with traditional models. The cost of compute wouldn’t make sense. But Vera Rubin drastically drops inferences costs I’m told. Taalas has an ASIC that can run inference that is 100 to 1000 times faster (read cheaper) than current generation GPUs. Maybe we don’t need 10,000 paths, maybe 100 will do. Maybe this isn’t so mad. Even if it’s just for a subset of limited used cases like real time risk.

The next problem of course becomes that not everything has a simple automatable verification. For some reason this isn’t getting talked about much and when I do bring it up, people have this perplexed look on their face. Like I’m the one that’s mad. Some complex derivatives can take literal hours of expensive compute to price (though admittedly we seem to trade less and less of these weird things). Validating each fanciful idea from an LLM to see if it improved things would cost a fortune even if the inference is cheap.

Peter Higgs proposed the existence of the Higgs Boson in 1964. It took 40 years and the construction of the Large Hadron Collider at CERN to validate that idea. 

MLCommons dropped version 6 of their benchmarks for AI inference last week amongst increasing concern for how we’re going to power all the GPUs that Nvidia is cranking out.  That’s probably just as well, given we’ll have to figure out which GPUs are going to be worth powering up and which ones we just leave sat in their boxes. Though as any seasoned HPC professionals will know, we probably should have started with the benchmarking and optimisation rather than the grandiose build plans. 

Whilst optimisation doesn’t give you the numbers you need to win the billionaire *cough* rocket *cough* measuring contest it might just give you a path to usable AI inference at scale. Might. If you’re running converged classical and AI models the latest MLPerf Inference v6.0 benchmarks are probably quite handy. More details and the results down below. Also HPC Wire took this moment to remind us that its system performance that will count. 

Most readers of this were already aware that we’re never going to be able to power all the data centres that have been provisioned, planned or promised. I guess current affairs means that the rest of the news cycle is catching up to that reality. Microsoft signed a deal with Chevron and that’s not something a younger version of me ever thought I’d hear myself say. And Google has decided that natural gas is the way forward to power AI. Maybe they didn’t get the memo about what’s going on in the world. 

Against this backdrop of news, France decided it was time to join the Exascale compute club. Not sure that means anything anymore with the HPC Top 500 being less and less relevant, missing as it is anything from the hyperscalers. I guess it still means something in the scientific community. Or in the measuring contests for politicians rather than billionaires. 

That was a rather cynical take this week. I better go outside and get some sun. And some Easter chocolate.

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 05 Apr 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. Log aggregation for Slurm on Azure and faster VM startup

Flux by HMx LabsHamza

MLCommons releases v6 of their inference benchmark. Results here

MLCommons Releases New MLPerf Inference v6.0 Benchmark Results - MLCommons

MLCommons releases MLPerf Inference v6.0 results — the most significant benchmark update to date, with new tests for text-to-video, GPT-OSS 120B, DLRMv3, vision-language models, and YOLOv11

MLCommonsMLCommons

Powering Data centres becomes critical

https://www.axios.com/2026/04/02/google-natural-gas-ai-power-energy

https://www.reuters.com/business/energy/microsoft-chevron-engine-no-1-sign-exclusive-deal-power-supply-2026-03-31/

We might not get (or need?) all the compute that was planned

Almost Half of US Data Centers That Were Supposed to Open This Year Slated to Be Canceled or Delayed

The vast majority of data centers scheduled for completion over the next few years have yet to even break ground.

FuturismJoe Wilkins

France set to join the Exascale club

Contract Signed for Alice Recoque, Europe’s New Exascale Supercomputer

The procurement contract for Alice Recoque, the new European exascale supercomputer to be located in France, has been signed by the European High Performance Computing Joint Undertaking (EuroHPC JU) and the selected vendor, Eviden.

The European High Performance Computing Joint Undertaking (EuroHPC JU)

System performance is what counts

https://www.hpcwire.com/2026/04/02/forget-about-chips-its-the-system-that-matters-for-ai/

From HMx Labs

 This was just an April Fool’s joke (though Corey Quinn says this entire category for faux series A pitch decks as a joke isn’t funny anymore).

HMx Labs Announces Continuum™

90% cheaper compute that feels like On Demand VMs but is built on unused spot capacity across multiple providers.

Flux by HMx LabsHamza

I hope you’re enjoying your Easter chocolate!

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

Neither I nor my AI bot could find anything from AWS this past week. Not even a new regional instance going live. I guess they’re too busy battling the fires Kiro lit.

Azure

Need a centralised way to look at your HPC logs? Want a turnkey solution for Cycle Cloud Workspace for Slurm? Azure has some ideas

Simplify troubleshooting at scale - Centralized Log Management for CycleCloud Workspace for Slurm | Microsoft Community Hub

Training large AI models on hundreds or thousands of nodes introduces a critical operational challenge: when a distributed job fails, quickly identifying the…

TECHCOMMUNITY.MICROSOFT.COMjesselopez

This could be handy, especially if building HPC capacity using spot compute. It's still in preview but we’ll definitely be playing with this to see the potential impact.

Azure updates | Microsoft Azure

Subscribe to Microsoft Azure today for service updates, all in one place. Check out the new Cloud Platform roadmap to see our latest product plans.

Microsoft Azure

Google Cloud

Nope. Nothing here either.

Spot compute, but it just… doesn’t stop.

Continuum is a deterministic execution layer built entirely on top of spare capacity across Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud, IBM Cloud, CoreWeave, Nebius and Nscale.

Yes, we know how that sounds.

Under the hood, workloads are continuously fragmented, moved, and rehydrated across a global pool of spot instances. You no longer need to care about which or how many CPUs or GPUs you need. 

No retries. No checkpointing ceremonies. No “this is fine” dashboards.

Think your workloads won’t work with this? Our Agentic Adaptation Tools say otherwise.

Just point the Continuum Adaptation Agent (CAA) at your repo.

It will:

• Refactor your code to run on Continuum
• Optimise for fragmentation, migration, and rehydration
• Quietly remove anything that assumes infrastructure behaves predictably

Not only that but you can also connect your on-prem cluster and sell spare capacity back into Continuum, turning that “idle HPC estate” into something your finance team suddenly cares about. And still use it in preference to renting someone else’s capacity.

Still stuck on CUDA and Nvidia? The CAA will refactor your code to use the Continuum GPU Rosetta Layer so you can use any available GPU or accelerator.

Continuum. Making CPU and GPU compute as easy to use as electricity.

As the dust settles on the hype created by GTC we saw that NemoClaw wasn’t meaningfully more secure than its ancestor OpenClaw and that the shine of growth forever in GPU clusters is starting to fade (in the financial markets at least). A week after HPSFCon though and we have hours of footage with actually useful information. Awesome.

Last week ARM told us that it will start creating its own CPU. Wait. Haven’t they already been doing that? No, they’ve just been designing chips for other people. So, you don’t run on an ARM CPU but an AWS Graviton or a Google Axion or possibly an Ampere. Now you’ll get to run on a chip that identifies not only its architecture as ARM but also its manufacturer in the lscpu output. It’s an interesting development and potentially puts ARM into competition with its own customers, a strategy that isn’t working out for Intel so well right now. I guess they’ve got Meta backing them at least though.

What I’m looking forward to though is to (hopefully) finally be able to test ARM against AMD EPYC and see how well the power efficiency claims hold up, not only when sat idle babysitting a GPU but also when running full tilt on some HPC workload. We already have the numbers on AMD, anyone want to give us access to some ARM kit?

At a time where people are questioning what happens to open source code in the face of the AI assault, IBM has just donated LLM-D to the Cloud Native Compute Foundation (CNCF) 🎉. Now if only they would open source their good HPC schedulers too and we could build a decent inference stack on open source IBM software.

Does anyone know what happened to insidehpc.com? All I get is a 404 error from WP Engine. It’s been like that for a couple of weeks now.

Aside from making Lego jokes last week we’ve also been busy giving you a way to look up and compare the details, performance and cost of various cloud VMs. Whilst continuing to vibe code (I mean agentically engineer) a new HPC workload manager and scheduler for giggles. Or more accurately, fail to vibe code one. More details on that down below as are all the links for the news above.

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 29 Mar 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. AWS gives us updates to both PCS and the other one (Parallel Cluster). Learn how to use TPUs from Google.

Flux by HMx LabsHamza

HPSF have made the videos from HPSFCon available

ARM gets into the CPU game directly

Arm Comes Full Circle With Homegrown, AI-Tuned Server CPU

It has been nearly five decades since British workstation maker Acorn Computer was founded, and near…

nextplatformTimothy Prickett Morgan

Meta Partners With Arm to Develop New Class of Data Center Silicon

We’re partnering with Arm to develop a new class of CPUs purpose-built to support data centers and large-scale AI deployments.

Meta Newsroomheathera

LLM-D goes open source. CNCF Style.

Donating llm-d to the Cloud Native Computing Foundation

llm-d offers a replicable blueprint for developers and researchers to deploy inference stacks for any model, on any accelerator, in any cloud.

IBM

From HMx Labs

AI is like a rusty old saw blade. Totally jagged

The Jagged Intelligence of AI

A non-update but still an update on HAL, our vibe coded HPC scheduler experiment

Flux by HMx LabsHamza

Want to know what’s in a cloud VM and how it perform and what it costs?

Announcing FLOPx: Data Discovery for Cloud VMs

Want to know what you’re getting when you spin up a cloud VM? How many cores, threads, memory? Actual benchmark performance numbers? We’ve got your back.

Flux by HMx LabsHamza

Make sure you keep your HPC sysadmin on side

And definitely play nice with your AWS account manager

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’m all for making HPC easier. Anything that does that is a noble effort. Even if it is the same technology that took down one of the most reliable online retailers in the world

Accelerating HPC Deployment with AWS Parallel Computing Service and Kiro CLI | Amazon Web Services

Research teams moving from on-premises HPC environments often struggle with the complexity of cloud deployment. Traditional approaches require deep expertise in AWS networking, storage architectures, and Slurm configuration management. A typical manual deployment involves weeks of infrastructure provisioning, network topology design, scheduler configuration, and performance tuning. Research teams with limited platform engineering resources find themselves […]

Amazon Web Services

AWS Parallel Cluster adds support for Slurm 25.11 and P6-B300 VMs

AWS ParallelCluster 3.15 with support for P6-B300 and Slurm 25.11 - AWS

Discover more about what’s new at AWS with AWS ParallelCluster 3.15 with support for P6-B300 and Slurm 25.11

Amazon Web Services, Inc.

AWS Parallel Compute Services also got some updates with the ability to set slurmbd and cgroup settings

AWS Parallel Computing Service supports slurmdbd and cgroups settings - AWS

Discover more about what’s new at AWS with AWS Parallel Computing Service supports slurmdbd and cgroups settings

Amazon Web Services, Inc.

New Regional Instances: R8gd in California, Seoul, Hong Kong, Jakarta, Cape Town and Calgary, M8a in Ireland

Azure

Must be tired from all the hype and recovery of GTC. Nothing going on here

Google Cloud

A guide to using TPUs for your AI training workloads instead of your usual CUDA

Training large models on Ironwood TPUs | Google Cloud Blog

Learn about the key techniques and tools within the JAX and MaxText ecosystem to improve training efficiency on Ironwood TPUs.

Google Cloud

AI can write code the same way that monkeys on a typewriter can produce the works of Shakespeare or the Monte Carlo method can calculate the value of Pi.

Ok that’s a little (maybe a lot) unfair. The AI generated output is not random, it’s the most probable output based on its training set but, and this is an important “but”, the validation required of that output is the same. In fact, the more complex the work or the more it deviates from the median of the training data of the model the truer this statement becomes.

And herein lies the problem.

It was trivially easy to generate an applicant tracking system or an company directory tool that does some basic audit tracking for internal use. Those are just CRUD operations with a pretty UI. The time taken to (create and) validate they operate correctly was less than the time it would have taken to learn how to use a SaaS offering.

Sadly, that translates very poorly to creating a new HPC scheduler / workload manager. Very poorly. Validating the behaviour of such a system is not trivial. That validation itself needs code and if that code too is AI generated, you’re in a circular loop of doom. 

This discrepancy is hard to talk about in a meaningful way. It would be very incorrect of me to just say AI can create CRUD apps but is useless in HPC. It would be a nice soundbite and provide a clear easy to follow bit of guidance on how to use a complex and novel bit of technology. But it would be wrong.

The answer is instead rather more frustrating. The capabilities are highly varied and for some reason also seem to vary through time for very similar tasks. I think the common phrase for this is jagged intelligence. The slightly shocking part is just how jagged it is, even within a small domain.

Ironically, part of the problem is also the very thing that we struggle with when trying to use natural language to program computers – the imprecision of language. We call anyone that works with code a programmer or software engineer or some other variant that is often used interchangeably. You don’t call yourself a carpenter if you put up an Ikea bookshelf. Yet we have exactly that level of disparity between what software engineers do. I’ve talked about this at more length in the past (even before AI) but I think it’s even more relevant now for AI adoption because not only does it differentiate between the capabilities of different people it also applies to different tasks the same person does as part of their job.

This has been a bit of a rambling update on the state of HAL and unfortunately the Git history won’t tell you a much better story because I got tired of the copy pasta dance of putting the AI history into Git and hitting token limits. I’ll try again and do better.

How many vCPUs did that 32xlarge AWS VM have again? Did the Azure Dav7 enable SMT by default or not? That Google N4D, that was that an ARM or x86 machine again? Oh, and how well do any of them perform for my own workload?

That’s always been the key question. And it’s frustratingly difficult to answer. We’ve been helping our clients answer it for several years and occasionally we’ll publish some public benchmark data, but I figured we’d step it up a notch.

Want to see the full CPU specs of the machine, not just how many CPUs but if SMT is enabled (and if you can disable it)? lscpu output? Performance benchmark data? Pricing and performance adjusted pricing? Now you can.

We quietly put FLOPx live a week or two ago in a pretty beta state but we figured we’d gather some early feedback. This is definitely in - what the hyperscalers would call -  Preview. We’re also limiting access to clients and partners (and friends of course!) for now (maybe forever?).

There’s an API to go with the UI for integration into your automated scaling mechanisms too. It’s time to take the guesswork out of HPC on the cloud.

If you’d like to take a look, give me a shout and I can enable your account for access in return for some feedback.

You’d be forgiven for thinking that the only HPC related news last week came out of Nvidia GTC. But you’d be wrong. Whilst GTC completely overshadowed it, last week we also had the second HPSFCon (High Performance Software Foundation Conference). I swear Nvidia does this on purpose. Last year GTC clashed with the International Supercomputing Conference. What’s up with that Jensen? Not cool man. 

Although it remains a little DOE centric still, and nobody can escape the monster that is AI, at least HPSFCon attempted to showcase some actual real uses of AI in HPC and scientific research. The levels of hype are also measurably lower! There isn’t a great deal of content available from HPSFCon just yet but there is a good video on converged computing and agentic AI in scientific research including using agentic AI to run and control your HPC jobs. If you’re already familiar with the concepts of converged computing and Flux framework, skip to 18:54.

Jensen pulled out faster inferencing from his hat at GTC with the Groq acquisition being put to use. He also showed of a pipeline of $1Trillion in sales though the financial markets don’t seem too convinced. 

Nvidia is getting a lot of heat on DLSS, and I’m not a gamer so I won’t comment on DLSS directly, but Jensen’s presentation mentioned how DLSS is the first application of using a combination of a physics model combined with a stochastic generative AI to produce something structured. I understand its  application in video games has done down like a seg fault in a MPI job just before completion, but that sounds kinda exciting. It aligns with the idea of combining genAI with traditional quantitative finance model and if things like Vera-Rubin/Groq/ ASICs really can drop the cost of inference sufficiently this space starts to look genuinely interesting.

Then again, it’s probably not that at all and just a bunch more random matrix math being hyped as sentience yet again and the gamers are probably right. 

Oh, Microsoft Azure announced a Blackwell RTX Pro powered VM too.

Anyone get any decent swag at GTC?

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 22 Mar 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. Azure gives us a new Blackwell powered VM and announces a lot of end-of-life dates. AWS makes it easier to share your capacity reservations.

Flux by HMx LabsHamza

HPSFCon: Converged Computing and Agentic AI in HPC. Skip to 18:54 for agentic HPC

From GTC

https://www.reuters.com/business/retail-consumer/nvidia-sell-1-million-chips-amazon-by-end-2027-cloud-deal-2026-03-19/

If you need real FP64 performance AMD might still have you covered even if Nvidia’s faith is the Ozaki scheme is misplaced

https://www.hpcwire.com/2026/03/13/amd-hints-at-big-fp64-increases-in-mi430x-gpu-as-ozaki-underwelms/

From HMx Labs

I slacked off a bit last week, so not a great deal here except to say Happy Eid

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 can now dedicate compute for HPC but still share it with other workloads (or between HPC workloads) when using capacity reservations

Amazon EC2 Fleet now supports interruptible Capacity Reservations - AWS

Discover more about what’s new at AWS with Amazon EC2 Fleet now supports interruptible Capacity Reservations

Amazon Web Services, Inc.

Enable CloudWatch across the whole organisations’ EC2 instances. Mind the bill

Amazon CloudWatch introduces organization-wide EC2 detailed monitoring enablement - AWS

Discover more about what’s new at AWS with Amazon CloudWatch introduces organization-wide EC2 detailed monitoring enablement

Amazon Web Services, Inc.

New Regional Instances: C8gn in Jakarta, Hyderabad, Tokyo, Sao Paulo and Zurich, M6in and M6idn (wow that’s old) now available in London, C8a in Tokyo.

Azure

New NCv6 VM family are now available in preview. Intel Xeon 6 CPUs with Nvidia RTX Pro 6000 (Blackwell) GPUs. 

Azure NCv6 Virtual Machines: Enhancements and GA Transition | Microsoft Community Hub

NCv6 Virtual Machines are Azure’s flexible, next generation platform enabling both leading-edge graphics and generative AI compute workloads. Featuring…

TECHCOMMUNITY.MICROSOFT.COMFernando_Aznar

If you’re still using them, we need words! But the HBv2 will finally be retired on 31 May 2027. Yes in typical Microsoft fashion you have over a year to deal with this. Same goes for the HCrs but I can totally understand why you’d be using those still as they can be rather good on the price/performance scale, especially for non MPI workloads

Azure updates | Microsoft Azure

Subscribe to Microsoft Azure today for service updates, all in one place. Check out the new Cloud Platform roadmap to see our latest product plans.

Microsoft Azure

Azure updates | Microsoft Azure

Subscribe to Microsoft Azure today for service updates, all in one place. Check out the new Cloud Platform roadmap to see our latest product plans.

Microsoft Azure

Azure Batch will finally retire using Low Priority VMs (if you even still remember what those are in) in favour of Spot VMs

Azure updates | Microsoft Azure

Subscribe to Microsoft Azure today for service updates, all in one place. Check out the new Cloud Platform roadmap to see our latest product plans.

Microsoft Azure

NVv3 and NVv4 EOL dates announced for use in Azure Batch

Azure updates | Microsoft Azure

Subscribe to Microsoft Azure today for service updates, all in one place. Check out the new Cloud Platform roadmap to see our latest product plans.

Microsoft Azure

Google Cloud

Nothing. Even in a week that included Nvidia’s GTC.

Its GTC this week, so with the best will in the world and as much as I’d like to focus on HPC any news is going to be focused around Nvidia and the AI monster. This kicked off already with Nvidia inflating its circular financial model with investment into Nebius, though it’s a relatively modest $2billion. The number go up analysts are already expecting Nvidia to announce further developments to improve its inference capability in a market that is rapidly providing other options. I guess we will have to see what uncle Jensen is able to pull out of his hat post the acquisition of Groq. All this before Vera Rubin has even hit the racks.

IBM gave us a reference architecture for quantum in supercomputing. I’m not sure how many are in a position to run production grade workloads on a quantum machine yet, but I guess it’s good to be ahead of the curve here. 

In case your budgets were not stretched thin enough by the jump in RAM prices due to the AI monster devouring everything it can get its hands on, CPU prices are now headed in the same direction as the beast tries to quench its insatiable appetite for compute. You didn’t want to buy any new hardware this year, did you? Beowulf clusters and running systems beyond EOL here we come! 🏴‍☠️

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 15 Mar 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. Some instance types available in new regions on AWS, some details on how to do air gapped inference on Azure and that’s your lot.

Flux by HMx LabsHamza

Will Nvidia pull a rabbit out its hat at GTC?

https://www.reuters.com/technology/nvidia-focus-competition-beating-ai-advances-megaconference-2026-03-13/?utm_source=chatgpt.com

IBM give us a reference architecture for quantum supercomputing

Unveiling the first reference architecture for quantum-centric supercomputing

At IBM Research, we’re inventing what’s next in AI, quantum computing, and hybrid cloud to shape the world ahead.

IBM

Nvidia inflates the circular economy with billions invested into Nebius 

https://www.reuters.com/technology/nvidia-invest-2-billion-ai-cloud-firm-nebius-2026-03-11/?utm_source=chatgpt.com

The AI hype monster starts devouring CPUs too

https://www.hpcwire.com/2026/03/09/ai-boom-comes-for-cpus-which-are-cool-again/?utm_source=chatgpt.com

From HMx Labs

We're trying to get HPC Club up and running for this year? Got any ideas?

Where’s HPC Club in 2026?

Have an opinion on what we should cover in our next HPC Club? Let me know!

Flux by HMx LabsHamza

HAL ran out of tokens!

HAL Update: No More Tokens

Not enough tokens to provide a real update but AI does seem to be struggling now

Flux by HMx LabsHamza

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

New Regional Instance: R8a in Tokyo, M8azn in Ohio, HPC8a in Tokyo, M8i and flex in Ireland and London, C8id in Spain, [CM]8gd in Sao Paulo and Ireland, R7gd in Sao Paulo.

Azure

I guess AI inference is a type of HPC if you’re serving enough users. Doing it in air gapped environments on cloud? Didn’t think that was possible and I guess if you use the strict sense of the word air gapped it’s not but this is what you can do in Azure. 

https://techcommunity.microsoft.com/blog/azurehighperformancecomputingblog/ai-inferencing-in-air-gapped-environments/4498594

Google Cloud

Not much going on here 

I was hoping to provide an update today that said the caching component of HAL (the AI first coded HPC workload manager/ scheduler) is done. I can’t. I ran out of tokens.

I asked Codex to create the actual implementation, which it kind of did…with kind of doing the heavy lifting in that sentence. When I asked it to actually run the tests and make sure they pass it seemed to just get into a bit of a funk and span it wheels for ages and then spat out:

“Selected model is out of usage You've hit your usage limit for GPT-5.3-Codex-Spark.”

All that trying to fix a bash script (which it wrote):

hamza@devbox tests % ./run-full-cache-tests.sh.
/run-full-cache-tests.sh: line 218: unexpected EOF while looking for matching `"'

🤣

Oh well. I have both a Copilot and a Claude subscription I guess I could point at it I suppose. Or I might just wait.

I have to say though, the implementation it has generated feels a little sparse. Either this thing is amazing and it can reimplement ValKey++ in about one tenth as much code or (more likely) it just doesn’t actually do most of what is needed.

Might need to take a stap back and generate smaller work items for it and then set it off to work on each item (with less context in each one) in turn. I think I was perhaps being a little ambitious in attempting to one shot the implementation. Even with all the specification and tests in place.

I do really wonder about some of the claims out in the wild about what people have created with AI at this point though. Maybe Claude’s uptime metrics and bug count are better indicators of AI’s true capability.

GitHub - hmxlabs/hal

Contribute to hmxlabs/hal development by creating an account on GitHub.

GitHubhmxlabs

If you’re wondering what happened to HPC Club in 2026 you’re not alone. A busy end to 2025 and even busier start to 2026 has unfortunately meant HPC Club has taken a little bit of a back seat. Sorry.

We’re trying to fix that now and are aiming to do the next event just after the Easter break. I am a little stuck though on a couple of things. First up: What topic should we cover this time and secondly, who should we invite as a guest speaker.

AI seems to have all but consumed HPC (in the media) and coming up with current talking points and areas of interest that aren’t AI related and probably already covered elsewhere is getting harder and harder.

I do have some possible suggestions. Firstly, hybrid workloads that contain elements of both LLMs and traditional scientific/ analytics workloads are starting to rear their head. In the traditional HPC space this is not too much of a big deal. What I see as being more problematic is where HTC workloads are converging with LLMs. I think that’s going to be an interesting problem.

The second idea is the rising interest in compute exchanges driven by demand (and/or surplus) GPU capacity. It’s something I’ve been talking about for close to a decade at this point but could AI be the driver that pushes the into existence?

Those are just my thoughts though. I want to hear yours. What do you want to talk about at HPC Club? Who do you want to hear from?