2 ms·
...Not precisely. AMD implements the "AI" matrix instructions in the shaders themselves, not as big seperate blocks like Nvidia. But unlike Nvidia, the instruct
by brucethemoose2 3y ago
...Not precisely. AMD implements the "AI" matrix instructions in the shaders themselves, not as big seperate blocks like Nvidia. But unlike Nvidia, the instructions are different on the consumer (RDNA3) and server line. I dont know anything about building rocm, but supporting rdna must indeed make things more difficult.
Intel takes this approach too.