How should the public MI308X versus H20 comparison be read?
A third-party public comparison (Moreh / Step3) reports MI308X decode at about 1.30× H20 and prefill at about 0.70× H20. That is a decode and long-context positioning. It is not our signed measurement, and it does not say the whole stack is faster.
- Our own storage figure is 29.15 seconds to 0.37 seconds versus local NVMe. Do not merge the two comparisons into one sentence.