Reverse-engineer the MME on pre-Turing
Basically, we need a repeat of !97 (closed) for pre-Turing GPUs, including a simulator, unit tests, and builder support. Currently it's TBD just how much we can share with the builders. We may end up needing different builders entirely but I'd like to avoid that is possible. This is the number one blocker for pre-Turing GPU support.
Edited by Faith Ekstrand