Huawei Ascend 950 reportedly adds on-chip routing without using DRAM bandwidth
bookwormengr · x · 2026-07-23
Huawei’s Ascend 950 NPUs appear to support an UB-Mesh style setup where chips can act as routers as well as endpoints.
- The IO complex sits on separate dies connected to the main dies via die-to-die (D2D) links.
- Each side exposes 9 ports, and traffic can be routed from any port to any port on that side.
- Forwarded traffic stays entirely on the IO die, so it does not enter the compute die and does not consume DRAM bandwidth.
- The design allows mixed injected and forwarded traffic, which could make networking and service planning more flexible.
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 tokens per megawatt vs B200 — thursdai_pod · 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