(无标题)
The xAI 10T param model pre-training is taking ~2 months
based on they are using around 20-40k GB200
theoretical 5e26 to 1e27 fp8 flops
with like 20-30% MFU that's ~ 1e26 - 3e26 flops
on top of that a low multiple (1x to 3x) for RL
so in total it's probably between 2e26 -