RightSignal
1 of 5

Your RightSignal

AMDAI serving

How AMD's ATOM engine serves coding agents on its MI355X chips

Coding agents re-send almost their whole history on every call, so a server that forgets that history must redo a lot of work.

Caching of history across GPU and CPU memory, a scheduler that groups short new requests so active agents keep generating, and fused decode kernels.

17.5K tok/sper chip with 56 concurrent sessions on eight GPUs
17 Nov 2026Read the post