Hacker News AIT2·72 pts
- Two tensor-parallel engines per trn2.48xlarge, each spanning 32 logical NeuronCores behind a session-affine router
- Decode runs as NKI megakernels, one fused kernel per layer half covering KDA attention, fused DSA, mHC and MoE FFN
- Prefix cache uses a radix tree over FP8 KV pages plus a host-memory tier for evicted pages and state checkpoints