Hey everyone
I was searching around if the Orin Ampere GPU supports Multi-instance GPU (MiG) technology, an eventually found in a presentation on the Hot Chips 34 conference in August 2022, that it is actually implemented in hardware to split into multiple instances:
This is all the information I could find surrounding the topic. Could someone provide some more information on this topic? Is it really supported, how many times can it be partitioned, can Triton utilise this… ?
If you want to check the presentation yourself: