What Makes HPC in Finance Different?

There’s no MPI. No one really cares about topological placement or colocation of workload. There are no high-speed interconnects like Infiniband. It even runs on boring commodity hardware. There isn’t even a parallel file system! I mean is it even really HPC?

I am of course talking about HPC in financial services. Talk to anyone within this space and your years of experience running HPC outside finance counts for nothing. Talk to anyone with a research computing background using supercomputers in the Top 500 and the above doesn’t even register as HPC. Yea… those are some fun conversations!

So, what gives? What’s the big deal here? 

Firstly, the absence of your typical supercomputer infrastructure like high speed interconnects isn’t an accident and I can assure you there are many times I wish we did have them! It is a deliberate choice to design a system that is able to run on highly commoditised hardware and scale horizontally instead of vertically.

This presents a completely different set of challenges in the software. Not only in the financial analytics libraries but also in the code responsible for enabling them to run at scale. This includes not only the risk system itself (more on that another day) but also the scheduler and other associated digital plumbing.

I guess for this to make any kind of sense, you need to have an idea of the scale a typical “too big to fail” bank will operate at. Let’s try and throw a few ballpark numbers out there. 

We’re talking in the region of 100,000 CPUs. Tens of thousands of compute tasks submit every second. Yes second. Your entire compute estate operating at close to 100% utilisation for hours on end. Then disappearing as you run out of work and tear it all down.

And all this has to happen day in, day out without fail, upon pain of huge fines or trading losses. (That last bit has a tendency to focus the mind).

I get it. It’s not HPC in its classical sense. Neither are the current crop of “supercomputers” being used for AI inference (and to a lesser extent to train AI models).

It’s a different approach to solving a very similar problem, enough compute to get your answer. 

Sometimes having another arrow in your quiver might just be what get you over the finish line.