FFN Splitting Accelerates on NVLink Racks
bookwormengr · x · 2026-07-14
This post discusses the concept of splitting Attention and FFN/MoE layers across different GPU servers in the latest rack-scale GPU systems.
The quote highlights an open-source implementation called FastAFD: an AFD runtime designed for GB200 NVL72 (72 Blackwell GPUs within a single NVLink domain). The author shares that it can boost decoding throughput per GPU by 1.35–1.45x. The post includes links to the code and a blog post, emphasizing this as a system-level optimization idea for next-gen GPU racks.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27