NVIDIA Made 4-Bit AI Work. It’s a Nightmare.

0:00 · Watch on YouTube

Nvidia’s NVFP4 is a microscaled 4-bit format plus a training recipe that tries to keep large-model quality close to FP8 or BF16 while new GPUs put their biggest throughput gains in FP4. The quoted 50 petaflops on Rubi...

Read the full summary on retube