What is the measured cold-recovery time versus local NVMe on MI308X?
The load is Qwen3-Coder-480B-FP8, TP=8, long-context cold recovery, compared with a local NVMe drive. Time to first token on a repeated prefix went from 29.15 seconds to 0.37 seconds. Source: reports R2/R3.
- Use these absolute seconds for the storage result. Do not recast them as a multiple versus another accelerator.