I have no idea of AMDs AI capabilities. How does a MI300x or a compare to a gh200? If it is even comparable. Does AMD have a GPU/Accelerator to train LLMs?
My mate in the LLM world borks at the idea of using AMD. All his workloads are done on nvidia. I can’t see that changing until AMD offer us something disruptive, which I doubt this is, sadly.
Even if they were matched, the answer is simple:
It’s because of CUDA.
I hope those dedicated AI cores on the 7000 series get put to good use with this upcoming announcement.
There are no dedicated AI cores in RDNA3 (RX 7000)
AMD specifically call them “AI accelerators”. They schedule the matrix computations to be calculated by the texture units on the card using WMMA instructions
This is the same technique they use for ray tracing. They don’t have dedicated cores. The ray bounces are traced by texture units. That’s why AMD is behind Nvidia in RT
It saves them money on die size
Texture units are fixed function.
I can understand the push for money, selling MI300s or whatever. What I don’t appreciate is the bs regarding improvements they say they’ve made with current drivers for consumer cards and the vapaorware (both Linux and AMD). I’ve had amd cards since the AIW cards from ATI, I’ll never buy another.
I kind of figured they’d wait for CES.
mi300 soft launch incoming
The fastest product in AMD history to rake in $1 billion in sales is hardly a soft launch…
AMD seem to have waited on MI300A to focus on AI