I’m a Go / Kubernetes engineer (7 years) working on LLM inference control-plane and scheduling for a platform that runs across heterogeneous domestic accelerators — Ascend, Cambricon, Hygon, Moore Threads, Kunlunxin — alongside NVIDIA.

Most public LLM-inference benchmarks only cover NVIDIA. I happen to have hands-on access to a rack of domestic accelerators, so I publish the head-to-head numbers almost no one else can — same model, same load, measured the same way — plus the routing, KV-cache, and cost trade-offs behind them.

My rule: every number ships with the exact command to reproduce it. If I can’t show you how to reproduce it, I don’t publish it.