AMD Strix Halo is an MoE part, not a dense-70B part - the ~256 GB/s wall caps dense decode at ~5 tok/s

A Strix Halo 128GB box bought to run a dense 70B model delivers only ~5 tok/s - far slower than the marketing 'runs 70B/120B' implies. The same box runs gpt-oss-120b MoE at ~34 tok/s, which is the number that gets quoted.

Token-generation speed ≈ memory bandwidth ÷ bytes-read-per-token. Strix Halo's one 256-bit LPDDR5X-8000 bus gives ~256 GB/s (200-270 effective). A DENSE 70B reads ~40 GB/token -> ~5-6 tok/s. An MoE like gpt-oss-120b reads only the active experts' bytes (~5.1B of 117B) -> ~34 tok/s. The wall is bandwidth, and it is physics, not a software fix.

Match the workload to the hardware: buy Strix Halo for gpt-oss-120b, GLM-4.5-Air, Qwen3-30B/235B-A3B-class MoE, or a silent always-on agent box - it is the best value here. For DENSE 70B buy bandwidth instead: Mac Ultra (~800 GB/s, ~12.6 t/s) > M4 Max (~546) > EPYC Genoa (~460) > Strix Halo (~256) > Milan (~205). A used M1 Ultra 128GB does ~12.6 t/s on the same dense 70B for similar money.

This is the single most-misread spec in the big-memory slice: 'runs 120B' is true but only because MoE is bandwidth-cheap. Measured (lhl/strix-halo-testing): gpt-oss-120b MXFP4 33.7 t/s, GLM-4.5-Air 23.4, Qwen3-235B-A22B 15.1, but dense Llama-3.3-70B Q4_K_M ~5.0 and dense Qwen3-32B Q8_0 6.4.