Heterogeneous Inference, Measured

Go / Kubernetes engineer (7y) working on LLM inference control-plane & scheduling across heterogeneous domestic accelerators.

Here I publish reproducible benchmarks — the same model, same load, run head-to-head on NVIDIA vs Ascend / Cambricon / Hygon / MetaX / Kunlunxin — plus the routing, KV-cache and cost trade-offs behind them.

No vibes. Every number ships with the exact command to reproduce it.