Slurm Workload Manager

Did it warrant that sort of rating? And he gave, you know, an answer that will surprise many of you.

I think in terms of what ClusterMax is aimed at, I mean, what it does is it rates managed Slurm and Kubernetes clusters.

We don't have any live managed services today.

And so we did not and could not provide them clusters to test.

Things like serverless inference endpoints, posted like post-training infrastructure for RL, harnesses, sandbox, you've been digging in there.

When you're talking to the providers themselves about the roadmap, do you find it confusing that people can't get Slurm and Kubernetes right sometimes, but they're already ready to launch like four new products?

I mean, I guess, right? Because...

Currently, the margins on compute are so high.

It's interesting.

As Prat said, there is an

interesting fluidity to the market right now where some of the better Managed cluster providers know how to manage a cluster because they have teams internally who are trying to use these things and can give good feedback for what a good Slurm on Kubernetes layer looks like, how to set up health checks, that sort of thing.

So there definitely is a benefit to integrating these things in-house, in addition to just the fact that you have another service that you're offering on top where you can take margin from another layer.

The only intervention that removes this bottleneck is for a funder to establish, or back, an entity that owns compute contracts and allocates GPU time across their grintease.

This might look like a standing allocation of a few thousand B300s on a hyperscaler or top-tier neocloud, for example CoreWeave or Nebius, managed as a shared cluster in the manner of a SLURM-based academic facility, with organizations allocated varying amounts of compute per project or fixed term, for example, six months, renewable.

This is the fastest way to translate capital into GPU hours.

Procurement is entirely amortized, with capacity risk managed centrally so that independent research organizations have reliable access to compute and can scale without re-entering the market.

So if you really look at the pieces themselves, so what do you have? You have compute, GPU, you have networking, you, you have ideally, um, fast storage, uh, that allow large leaning, um, models or LLMs, uh, pipelines, and again, data inferencing.

Now, what-- when you talk about scheduling, uh, these jobs or processes etc., you have Slurm, uh, or, uh, you could have, uh, Kubernetes.

Um, and then, and beyond that, you actually have an NVIDIA microservices or CUDA libraries as well.

So it's not just as simple as it used to be when, when you had a small number of people and, uh, really a small degree of expertise.

The second example I can call out, which one of our customer face is in the AI cluster front, right? So the flow stalls, thousands of GPUs are there, the flows are stalling and creating problems with the training.

How do you go about it? How do you go about troubleshooting these issues? So we have, for example, integration with job schedulers like Slurm, where we understand the training, we understand how the training is distributed in the network, and then we understand the flows in the network flows as well.

So with that, if there is a job training causing an issue, you can simply ask, hey, why is my job running slow? Or you can go to our system to see which job has impacted with which congestion at what point in time.

That's key, right? Because it's sporadic.