Kimi K2.6 hits 264 t/s on MI50s

This is a strong systems benchmark, but it reads more like a proof-of-concept for keeping old Instinct hardware relevant than a practical deployment recipe.

• –The result is split between generation throughput and prefill throughput, so it favors batch ingest and long-context work more than low-latency chat.
• –PCIe instability and only partial bandwidth on the risers likely left performance on the table; the author explicitly expects better numbers with cleaner links.
• –The real story is the AMD ROCm/vLLM fork: `vllm-gfx906-mobydick` is doing the heavy lifting to make gfx906-era cards useful for modern MoE models.
• –Kimi K2.6 is a sensible stress test here because it ships as an open-weight model with native INT4 support and is meant to run on vLLM-class stacks.
• –For most teams, this is a curiosity unless power is cheap; for tinkerers with used MI50s, it is a useful datapoint on the ceiling of self-hosted inference.