Is there any way to know the approximate cache evict policy for A100/H100? Do nvidia have document about this? Or if someone have do some research about this?
Is there anyway to help to select data to load in the L2 cache
You can use L2 carveout on A100/H100.
Yes, but I just want to know the default policy for evict, ot just the simulation of this
Yes, I was mostly responding to this: