Magical Cloud VMs
Can running HPC on the cloud actually result in higher performance than bare metal on prem? Turns out it can!
Given the fastest supercomputers in the world are owned by hyperscalers it seems a little strange to even to have to state that it’s possible to do high performance computing in the cloud.
Even so, HPC on the cloud is often still seen as a bit of a (necessary) compromise. The idea that some performance is being left on the table to pay for the hypervisor lingers, and bugs, many practitioners.
But what if that wasn’t true?
During our testing of AMD’s 4th and 5th generation CPUs we thought it might be fun to try and see what that performance penalty actually is. Imagine my surprise then when our results showed that not only is there no performance penalty but that the cloud VMs score higher than the bare metal system we were testing!
As clever as all the wonderful people working at the hyperscalers are, they still have to obey the laws of the universe and can’t (yet) just magic up benchmark results. They can’t, right? Right?
So, what gives? How is a virtual machine faster than bare metal? When you buy hardware at the scale hyperscalers do, you don’t pick CPUs out of the catalogue. You get to specify what you want. Unfortunately, they don’t tell us the exact specs of the CPUs they have. At a guess I’d say they have higher clock speeds or larger L3 cache sizes. Or maybe both.
Whilst this is not an exact apples for apples comparison (because the hyperscalers use a custom model of CPU), it does reflect the reality of options available to most of us.
I think I’d like to get our hands on the entire lineup of 5th generation CPUs next… curious minds have more questions!
See below for the full paper
