Does using Kubernetes to run HPC workloads really make sense?

I have an opinion that I don’t usually share online for fear of being ridiculed. But yesterday I shared it with a stranger and lived. Plus, its Friday so even if I’m totally wrong on this everyone will have forgotten by Monday anyway. I don’t think Kubernetes should be used to run HPC workloads.

I’ll admit, it has been a number of years now since I last did this in anger at any kind of scale and I’m certain that K8 has evolved considerably since then. We’ve definitely had things like Slinky released to make it easier too. Even so, I don’t see a compelling reason for the additional complexity a K8 cluster brings when all you really care about is access to compute. It only gets worse when you have to pass through any kind of specialised hardware.

I think at some point in time, it made sense because K8 became a defacto way to access cloud capacity in something close to a vendor agnostic way. I get that. I don’t think it’s a good reason anymore. While the clouds may not have a common API there is more than one commercial product and open source project that let you work around those limitations. Even without those, having to deal with two or three APIs to manage resources is really not that much work. Especially in an era where your favourite genie will whip that up in the blink of an eye.

Conversely, I think from a cloud provider’s point of view, I wonder if it makes it easier to then pull you along the path to additional cloud native products. And we all know cloud native was just a great marketing term for cloud locked right? 😁

Change my mind.

For roughly two years I’ve been putting together release notes from the major cloud providers that are relevant to people that care about HPC. By Hand. Like some kind of caveman. Every week I’d read the release notes and pick out the bits that are HPC relevant.

I did try and automate it a couple of times, for some definition of try. Yea ok I just told Claude and ChatGPT to do it in a web chat window. I didn’t try very hard. Anyway, that never seemed to work very well. Mostly I seem to struggle getting it to correctly identify what might be HPC relevant.

Last weekend though I tried a bit harder. I AI first engineered, nah who am I kidding, I straight up vibe coded a web app that reads the RSS feeds from the cloud providers, attempts to classify them, mostly fails and gives me a way to swipe left or right on each release note so you still get a nice, curated list.

It needs a little (maybe a lot?) more work before I put it live and I’m sort of playing with writing my own ML classifier based on the last two years of data rather than using an LLM.

Don’t worry, one of the items on the to-do list is a way for me to add the odd snarky comment against the releases like you get in todays version 😆

Now that this process is a bit more scale-able, not only can I add a more clouds and possibly neo clouds to the RSS list, but other vendors too that produce HPC relevant products. Well so long as they have an RSS feed, I guess. So, who else should I add? 

More importantly, it will need a name. Suggestions?

Cerberas decided to attempt another IPO last week while Bolt Graphics taped out its first GPU in an attempt to reclaim some of the vacant space in the FP64 space left by Nvidia’s all while Google also announced its 8^thgeneration TPUs. The market for accelerated computing certainly seems to be a lot diverse than it ever has been in the past and it will be interesting to see how many, if any, of these alternatives to Nvidia’s CUDA ecosystem manage to gain a significant foothold.

 I would say the choice is nice to have, but the reality is we can barely even swap across CPU architectures in many domains (more on this in a bit) let alone across the significantly more difficult accelerated computing domains. For now the choice may possibly give us a more competitive market (and even then I’m not so sure) but for most people it will still mean picking a direction and sticking with it for a number of years if not decades.

The timing of Cerberas’ IPO is also interesting, being as it would, the first major AI IPO this year. Are they trying to get out ahead of OpenAI and Anthropic who leave about as much liquidity as there is in the Sahara?

Meta announced a partnership with AWS and adoption of Graviton CPUs amidst increasing chatter that we’re running short of CPU capacity too now. Allegedly this is because agentic workflows demand more CPU capacity and are seeing GPU to CPU ratios collapse to close to 1:1 instead of the current 2:1 or even 4:1. I don’t buy it. At least not for CPUs collocated with GPUs. It simply makes no sense to burden a GPU machine with anything other GPU workloads. You’d architect any system at scale to offload CPU only compute elsewhere. And quite frankly I don’t buy that it’s even CPU bound at all. It’s more than likely IO bound. I’m sure Intel/AMD/ARM would love to make more money on their CPUs, and this is a useful narrative, but I don’t for a minute think it’s true. At least not for the reasons being proposed.

In our own news, we’ve moved the needle a little in making compute workloads more portable both across clouds and CPU architectures last week. Well technically it happened over a longer period, but we talked about it last week! Checkout FINOS ORB and our work on automating resolving numerical stability challenges when moving CPU architectures.

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 26 Apr 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. PCS supports Slurm 25.11, Azure gives you NetApp ransomware scanning and Google Cloud Next Updates.

Flux by HMx LabsHamza

Bolt Graphics provides an FP64 alternative to Nvidia GPUs

https://www.hpcwire.com/2026/04/22/bolt-graphics-targets-fp64-hpc-workloads-with-zeus-gpu/

Cerberas tries for IPO again. First of the AI IPOs this year?

The Second Time Will Be The IPO Charm For Cerebras

Waferscale chip pioneer and AI systems maker Cerebras Systems filed to go public back in September 2…

nextplatformTimothy Prickett Morgan

We need more CPUs for AI? I don’t buy it. This is a lazy analysis

CPU requirements for AI workloads are multiplying, driving intensifying shortages and price hikes — Intel already shifting production from consumer chips to Xeon as inference workloads drive server CPU ratios back toward parity with GPUs

Intel says CPU-to-GPU deployment ratios have tightened from 1:8 to 1:4, and could reach 1:1.

Tom's HardwareLuke James

Meta still on a spending spree for hardware, this time Graviton CPUs

Meta Partners With AWS on Graviton Chips to Power Agentic AI

We’re announcing an agreement with AWS to bring tens of millions of AWS Graviton cores into Meta’s compute portfolio to support agentic AI workloads.

Meta Newsroomheathera

From HMx Labs

The end game is to make HPC workloads more portable, so we had a couple of small wins in that space last week. Progress in running across different clouds and CPU architectures.

FINOS ORB: One Cloud API to Rule Them All?

FINOS announced the Open Resource Broker yesterday, an adaptive layer to allow HPC schedulers to control cloud capacity designed to work across multiple clouds and schedulers. We’re all on in and are helping to create it.

Flux by HMx LabsHamza

Testing for Numerical Stability and Can AI Help Fix It

Can LLMs fix numerical stability problems in your code. Let’s design a little experiment to find out and how much would you just like a way to just validate your numerical stability across CPU generations and architectures.

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

Sagemaker optimises your Slurm topology automatically now

Amazon SageMaker HyperPod now supports automatic Slurm topology management - AWS

Discover more about what’s new at AWS with Amazon SageMaker HyperPod now supports automatic Slurm topology management

Amazon Web Services, Inc.

PCS Support for Slurm 25.11

AWS Parallel Computing Service now supports Slurm 25.11 - AWS

Discover more about what’s new at AWS with AWS Parallel Computing Service now supports Slurm 25.11

Amazon Web Services, Inc.

Relying on AWS Compute Optimizer to do the heavy work figuring out if you’re using the right kind of EC2 instance? It now supports a lot ore instance types… but I have an even better answer for you

AWS Compute Optimizer supports 162 new EC2 instance types and 32 new RDS DB instance classes - AWS

Discover more about what’s new at AWS with AWS Compute Optimizer supports 162 new EC2 instance types and 32 new RDS DB instance classes

Amazon Web Services, Inc.

This sounds really promising but I do wonder how effective it is in reality. 

Amazon SageMaker AI launches optimized generative AI inference recommendations - AWS

Discover more about what’s new at AWS with Amazon SageMaker AI launches optimized generative AI inference recommendations

Amazon Web Services, Inc.

New Regional Instances: C8i-flex now in Ireland, London and New Zealand, C8i in Ireland and New Zealand, G7e in Los Angeles

Azure

If you’re using NetApp for your HPC file system you can now scan for ransomware to slow down your compute!

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

8^th Gen TPUs for your AI workloads

TPU 8t and TPU 8i technical deep dive | Google Cloud Blog

The 8th generation TPUs are engineered with system-level co-design to accelerate the AI lifecycle. TPU 8t is built for frontier-model training and TPU 8i is built for large-scale inference and reinforcement learning.

Google Cloud

Details on scaling compute (CPU and GPU)

What’s new in compute at Next ’26 | Google Cloud Blog

Learn how to scale web servers, databases, and enterprise applications to fuel your AI agents with the latest offers on Google Compute Engine.

Google Cloud

Google’s tips on cross cloud

Cross-cloud infrastructure at Next ’26 | Google Cloud Blog

Next ’26 cross-cloud infrastructure enhancements in four areas: fluid compute, connectivity, a unified data layer, and digital sovereignty.

Google Cloud

A recap of everything from Cloud Next. I had a scan through and there are a few that are HPC relevant but not new from last week.

Google Cloud Next 2026 Wrap Up | Google Cloud Blog

A whirlwind recap of Google Cloud Next 26, including a synopsis of over 250 product, customer and ecosystem announcements.

Google Cloud

I think I’m bored with trying to vibe code a HPC scheduler. I do have another, probably more useful and potentially more feasible use case for your little AI friend though. Numerical Instability.

I don’t think it’s possibly to specify in sufficient detail the test cases (even if you vibe code the tests themselves) when a system becomes as complex as the HPC compute and data scheduling system I was attempting. Without those test cases it’s impossible to have the validation step in the AI generation loop. Without that validation step you’ll never have output that works. Plus, I just keep running out of tokens! 

Skill issue? Maybe. Mythos required? We’ll see. I don’t think I’m on Anthropic’s special list though. Oh yea, all I need to do is change the model name in the API call anyway 🤣

I haven’t totally given up; I think I will come back to this when I have a little more time and perhaps money to blow on the experiment. Till then, I have something that might actually work.

If your little LLM pet can scan your code for security vulnerabilities, I wonder if it can also scan your code for numerical stability issues.

I’ve created a test case using QuantLib to price a digital barrier equity option using a Monte Carlo engine. It’s stable across multiple runs on the same machine (i.e. MC stable) but it does give me small but significant differences when run across different CPU architectures.

The next step is to create that all important automated verification. I have a couple of choices here and I’d like a little input from you. I could just have a couple of machines that I let run Claude run wild on with SSH access and see what it does and just compare numbers coming out the end. Cheap to do. Not very useful once it’s done. Or I could take the engine behind FLOPx that runs all our cloud benchmarks and change it every so slightly so that it can it runs a numerical stability test instead. That would give me a reusable, way to plug in any code to be run across multiple cloud VMs and then validate the answers from each are the same.

In the same way that we can now confidently and consistently benchmark cloud VMs at scale, this would allow us to validate numerical stability across multiple CPU generations and architectures. For the cost of running a few spot VMs.

In either case, I can then create a harness that attempts to find (and fix) the numerical stability problem – perhaps I can create a skill that allows it to use our Intel PIN tool extensions to help – using the aforementioned validation in a Karpathy style auto research loop and let it get on with it while I sleep. Might still need mountain of tokens though.

My question to you is: Do you care about having a numerical stability verification system? Or shall just write a bash script comparing two numbers and let Claude run rampant on two or three sandboxes.

Ever wanted a single unified interface to all the clouds? One API to rule them all? Yea me too. We just got a small step closer to that being reality.

Yesterday FINOS (part of the Linux Foundation, focused on financial services) announced the Open Resource Broker. An adapter designed to allow HPC schedulers (or really anything) to provision compute capacity on clouds. It has been contributed jointly by Morgan Stanley and AWS, thank you kindly both!

Before you roll over and conclude that this is just another AWS ploy to get you to use their cloud or even that this is only for Wall Street types, I want to say something that I hope will change your mind.

HMx Labs are all in on this too. In fact, we’ve already written most of the code needed so that ORB will not only work with AWS but Azure and Google Cloud too. From near enough day one. 

I also hope I’ve built enough credibility with you dear reader that despite having a finance background you’ll believe me when I say that there is nothing in here that prevents anyone, anywhere from using it with whichever HPC scheduler you want. Or None. Just ask for cloud capacity without a scheduler if you like.

This is the internet, so I know you’re going to ask for receipts. Give me shout. I’m happy to talk.

Major Banks, Cloud Providers, and Technology Leaders Announce Unified Open Source Fabric for Next-generation High-performance Computing (HPC) to Enable Faster Decisions and Breakthrough Efficiency

The FINOS community, including members Citi, Morgan Stanley, RBC, AWS, and Oracle, is advancing open HPC initiatives that deliver faster, smarter, more accessible, and drastically more efficient compute technologies for the financial services industry, through a unified, hybrid, and AI-ready fabric.

FINOSFINOS Team

It appears the US Government suddenly woke up about 4 months late and realised that Nvidia drinking all the Slurm might affect their national supercomputers. Senator Warren has written to the DOE about his but might not get the alarmed response she’s hoping for given that El-Capitan and most of the other largest supercomputers under the DOE now run on Flux Framework. Just as well really as they’re also based on AMD silicon. Coincidentally (or maybe not!?) Todd Gamblin from the Lawrence Livermore National Laboratory was kind enough to speak at the Google Cloud Advanced Computing Community last week too.

A lot of the news last week was around using AI to accelerate quantum computing. Nothing like powering your AI hype train with some quantum fuel right? And I think I’ll stop there before I say something that upsets people.

The slightly more interesting bit of news that caught my eye though was a piece from HPC Wire on data lakes for scientific data. The idea being to retain a lot more experimental output data than was traditionally ever done so that it can be used in AI models. I think the idea is actually broader than just the scientific community. In a world where potentially almost anything could form useful training data, retaining, categorising and cataloguing that data is suddenly useful. Important even.

This applies to so much that we traditionally ignored. Multiple categories of enterprise data, intermediate calculation results that were often used for debugging or compute optimisation but never kept long term (I’m thinking financial risk analytics here). Even telemetry data from production systems.

I think this is a space that hasn’t really been explored. HPC and AI data planes are pretty dumb still. They might be bigger and faster but they’re still not really any smarter. Getting the right data, from the right, authoritative source, from the nearest possible location. Yea that’s still not happening. Not even close. Agentic access to data is going to be pretty meaningless if the human still has to figure out where to get the data from. That’s 80% of the problem most of the time! 

Lastly the idea of sovereign HPC (funny how much more normal it seems when you don’t call it AI) hasn’t really gone anywhere. Both Canada and Euro HPC made a bit of noise and advances in this space last week.

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 19 Apr 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. Todd Gamblin of LLNL talks AI integration with the folk from Google. AWS embraces multi cloud. Azure lets you bring your own keys.

Flux by HMx LabsHamza

Did US Gov just notice Nvidia drinking all the Slurm?

https://www.reuters.com/technology/us-senator-warren-voices-concern-over-nvidias-acquisition-slurm-2026-04-15/

What’s the best thing to do to keep the AI hype train running, fuel it with quantum of course

Nvidia Brings The Power Of Open Source AI Models To Quantum Computing

For quantum computing to reach the point where it is fault-tolerant, scalable, and commercially viab…

nextplatformJeff Burt

What does data mean in an AI age? How much more do we need to keep? How much more meta data do we need for it to make sense? And how do we catalogue, categorise and manage all of it?

https://www.hpcwire.com/2026/04/17/the-rise-of-experimental-data-lakes/

Sovereign compute capacity hasn’t gone away

https://www.hpcwire.com/aiwire/2026/04/16/canada-opens-applications-for-ai-supercomputing-infrastructure-program/

https://www.hpcwire.com/2026/04/17/the-rise-of-experimental-data-lakes/

From HMx Labs

I’ve been a little too busy to write anything interesting and am also trying to focus on a couple of longer articles including on using AI within HPC and financial services. 

Here’s a little amusement to keep you occupied instead

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

Want multi cloud HPC? One step closer with AWS Interconnect now GA

AWS Interconnect is now generally available, with a new option to simplify last-mile connectivity | Amazon Web Services

Today, we’re announcing the general availability of AWS Interconnect – multicloud, a managed private connectivity service that connects your Amazon Virtual Private Cloud (Amazon VPC) directly to VPCs on other cloud providers. We’re also introducing AWS Interconnect – last mile, a new capability that simplifies how you establish high-speed, private connections to AWS from your […]

Amazon Web ServicesSébastien Stormacq

New network and block storage optimised cousins of the C8i are now available

Introducing Amazon EC2 C8in and C8ib instances - AWS

Discover more about what’s new at AWS with Introducing Amazon EC2 C8in and C8ib instances

Amazon Web Services, Inc.

Of course if you’re going to get that, then the Graviton versions need a performance boost to keep you interested

Amazon EC2 C8gn, M8gn, and R8gn instances now support higher Amazon EBS-optimized performance - AWS

Discover more about what’s new at AWS with Amazon EC2 C8gn, M8gn, and R8gn instances now support higher Amazon EBS-optimized performance

Amazon Web Services, Inc.

Azure

Users filling up their NetApp shares again? At least now you know who to blame

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

HBv2 is oooold. You really shouldn’t be using it anymore and know Azure won’t let you include them in Azure Batch Pools either

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

Bring your own key now works for Premium SSD v2 and Ultra Disks

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

If you’re trying to burst to cloud and using Azure Files then Azure File Sync might be handy and is available in a few more places now

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

Not much by way of updates but the Google Cloud Advanced Computing Community did manage to get Todd Gamblin (LLNL) over to talk about hardware diversification, AI integration and converged computing.

Nvidia dropped a technical paper last week delving into topology aware scheduling on rack scale GPU clusters. Subsequent to their acquisition of Slurm, which now seems truly integrated into Mission Control, I think this is the first in depth guide from Nvidia on how to do this. A sign of further investment to come into the Slurm ecosystem?

It came though, in the same week as a piece from Reuters, contributed to by Intersect360 on the hazards of Nvidia’s Slurm acquisition. Aside from highlighting that the concerns around this have not dissipated I don’t think it really said anything new or wasn’t said back in December last year at the time of the acquisition. In the intervening time though something else substantial has happened and I think worrying about the impact of Nvidia’s acquisition of Slurm is almost missing the wood for the trees. We are now in an environment where the use of AI in coding is having a massive impact on open source software and communities around it. Everything from maintainers drowning in AI Slop PRs to AI washing source code (to change its license) to simply more easily forking and maintaining your own versions of popular open source projects. (Not a fork, but extensions like S9S are a good example for Slurm).

While our own work confirms (at least to me) we aren’t going to see a vibe coded Slurm replacement any time soon, it also shows that forking Slurm and adapting it to your own specific use cases isn’t the challenge it once was either. My concerns around how Slurm evolves under Nvidia’s ownership remain, but honestly, I have other bigger concerns around FOSS more generally. There also isn’t a shortage of alternatives to Slurm.

In the classical HPC space the US DOE released SYNAPSE-I to handle real time data processing at HPC scale. On reading this I did wonder if they had spoken to CERN who face similar challenges, but I guess with geopolitics being the way they are right now that probably isn’t going to happen.

Our own news from last week: If you haven’t already told us, let me know what benchmarks you’d like to see in our data to make your infrastructure selection decisions. Also, the piece on LLM inference as Monte Carlo paths seems to have resonated with a few of you and is worth a read if you haven’t already.

Oh, and AWS gave us a real filesystem on top of S3. Given the number of HPC workloads already using S3 I give it all of 4 femtoseconds before the first HPC workload attempts to that too.

In The News

Updates from the big three clouds on all things HPC.

HPC Cloud Updates WE 12 Apr 2026

Updates to AWS, Azure & GCP in the last week relevant for HPC practitioners. A new Azure region, a filesystem over AWS S3 and running GPUs at scale from Google.

Flux by HMx LabsHamza

Running AI Workloads on Rack-Scale Supercomputers: From Hardware to Topology-Aware Scheduling | NVIDIA Technical Blog

The NVIDIA GB200 NVL72 and NVIDIA GB300 NVL72 systems, featuring NVIDIA Blackwell architecture, are rack-scale supercomputers. They’re designed with 18 tightly coupled compute trays…

NVIDIA Technical BlogRyan Prout

Reuters and Intersect360 worry about Nvdia's Slurm acquisition.

https://www.reuters.com/technology/nvidia-acquisition-schedmd-sparks-worry-among-ai-specialists-about-software-2026-04-06/

and our original written at the time:

Nvidia Drinks (Acquires) Slurm

Some thoughts and questions on what Nvidia’s acquisition of Slurm means for both AI and traditional HPC.

Flux by HMx LabsHamza

DOE goes real time with data analysis at HPC scale

https://www.hpcwire.com/2026/04/09/new-real-time-ai-system-closes-the-gap-between-data-and-discovery-at-doe-labs/

From HMx Labs

If using AI reliably means gating its output on automated validation, are we just aiming for LLM inference as Monte Carlo paths?

Is AI actually intelligent — or just a smarter dice roll?

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

Flux by HMx LabsHamza

Want to help drive the direction of FLOPx and make sure we have benchmark data that you care about? 

Hardware Selection Benchmarks

What benchmarks should we add next to FLOPx?

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

Perhaps not strictly HPC but I’m already aware of people using S3 (and other BLOB storage exposing an S3 API) in HPC workloads that I’m fairly sure that won’t be long before we see HPC running on top of this too

Launching S3 Files, making S3 buckets accessible as file systems | Amazon Web Services

Amazon S3 Files makes S3 buckets accessible as high-performance file systems on AWS compute resources, eliminating the tradeoff between object storage benefits and interactive file capabilities while enabling seamless data sharing with ~1ms latencies.

Amazon Web ServicesSébastien Stormacq

Maybe a little more HPC. NetApp is available in London, Hyderabad and Sao Paolo now

Second-generation Amazon FSx for NetApp ONTAP is now available in four additional AWS commercial and AWS GovCloud (US) Regions - AWS

Discover more about what’s new at AWS with Second-generation Amazon FSx for NetApp ONTAP is now available in four additional AWS commercial and AWS GovCloud (US) Regions

Amazon Web Services, Inc.

This might help allocate costs to your HPC clients if you’re running a platform with multiple users

Amazon EC2 Capacity Manager now supports tag-based dimensions - AWS

Discover more about what’s new at AWS with Amazon EC2 Capacity Manager now supports tag-based dimensions

Amazon Web Services, Inc.

New Regional Instances:  X8i in Paris

Azure

New Azure region in Denmark

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

I guess if you’re trying to run LLM inference using K8 this could be relevant. 

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

Google gave us a guide on building GPU infrastructure

A guide to architecting reliable GPU infrastructure | Google Cloud Blog

Understand the core challenges behind reliable GPU-based infrastructure on Google Cloud, and find resources to dive deeper.

Google Cloud

but also claims of better carbon efficient of its Ironwood TPUs

Ironwood TPUs deliver 3.7x carbon efficiency gains | Google Cloud Blog

The seventh-generation Ironwood TPU achieved an approximately 3.7x improvement in Compute Carbon Intensity (CCI) compared to TPU v5p architecture.

Google Cloud

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.

For the avoidance of any doubt: This was an April Fool's joke.

I had a few conversations where I had to clarify that. I know it sounds real and we do some work in parts of this space but we don't have Continuum as a product I'm afraid!