Comments by "kazedcat" (@kazedcat) on "AMD's CEO Wants to Chip Away at Nvidia's Lead | The Circuit with Emily Chang" video.

CUDA is the least of their problems. Their biggest problem is that everyone is moving to ASIC for AI. ASIC are inherently more energy efficient than GPU's. And power cost has become an enormous factor in developing frontier AI.
3
@foxale08 AI is closer to cryptomining than graphics processing. With AI you have one function that is used a lot which is matrix multiply. Cryptomining also has one function that is used a lot. The difference is that AI needs a tiny bit of general processing while Cryptomining can be all fixed function. The main difference between GPU and AI is memory access. Memory access for AI is a lot more structured so you don't need expensive predictived Cache Prefetcher that is used to manage GPU caches. In fact AI needs only a direct SRAM buffer instead of GPU's complicated multi level cache hierarchy and all the logic needed to manage this caches.
1