This is something NVIDIA gets right. Meanwhile, ROCm isn’t even supported on certain lower end AMD consumer chips.
Which is a shame because the Radeon VII is still better than anything else when it comes to 64bit float computation.
I watched geohot streams and if he can go for an hour with a need to restart the computer due to crashed kernel driver/locked up card, it's a miracle (various computers, one or multi-card setup).
And he is not trying to crash the driver, just making the most basic stuff work.
I seriously think there is a deep HW bug that caused the difference in performance between RDNA3 presentation and reality and they had to do some horrible hacks to make it work.
Either that or the card never seen a tester for compute. Which is strange, considering they added it as a "supported" option.
He's trying to do everything on AMD's consumer based GPU's, while he's comparing it with NVidia's enterprise GPU's with the same expectations.
That's just not how it works in real life. But if he succeeds, yeah, he could save tons of money and his hardware for ML could actually be completed.