엔비디아 블랙웰의 비밀
백색나무 · 51:13
The video argues that the expected split, with big general GPUs for training and cheap specialized ASICs for inference, has turned around. Sparse Mixture-of-Experts (MoE) models are cheaper to run per token, but every...
백색나무 · 51:13
The video argues that the expected split, with big general GPUs for training and cheap specialized ASICs for inference, has turned around. Sparse Mixture-of-Experts (MoE) models are cheaper to run per token, but every...
Redirecting...