deepseek-v4.1IMAGE
DeepSeek V4.1 Flash on AMD MI325X: 2.27x faster decode at 436 tok/s per user
Changing only the decode kernel with megakernels on 4x MI325X raised DeepSeek V4.1 Flash-1017 P90 output from 192.5 to 436.3 tok/s per user with no accuracy loss.
Sciforium
@sciforiumArchived 7 days agoOriginally posted 2026.10.24
Sciforium @sciforiumOriginally posted 2026.10.24
DeepSeek-V4.1-Flash-1017 on AMD Instinct GPUs: 2.27× faster code generation per user with megakernels, no accuracy loss Same weights, only the decode kernel changed: 192.5 → 436.3 P90 output tok/s per user on 4× MI325X. How we verify it:
x.com/sciforium/status/2113835976258990380Archived 2026.10.24 12:06