Comment by awesomeMilou

One thing that comes to mind is latency in a HPC context.

If you can fit your entire software stack into L3 cache, you never have to access memory and benefit from insanely good latency.

This was a thing in HFT like ~5-7 years ago, where you might have bought a 28 core Xeon, disable all but one or two cores and benefit from ~30mb L3 cache, which can fit an entire Linux Kernel, an SSH server and your trading software.