we should have the notion of "RGU" described somewhere in our documentation

[gyom]

We've been using "RGU" internally at Mila in order to sync up with DRAC, but it doesn't seem like the concept is explained anywhere in our documentation.

We should have a simple explanation of the notion of "RGU" used by DRAC as well as a table that lists the equivalences for popular GPUs. I want to be able to point Cursor at docs.mila.quebec, and get correct answers for the following questions:

• what's an RGU?
• why using RGUs instead of GPUs?
• how many RGUs is a H100?

[btravouillon]

I guess we should first agree with DRAC. RGU is their own term. If our computation is different, I'm not sure we should reference RGU on the Internet.