Strix Halo / CachyOS

The “fast local” tier — a big-iGPU desktop-class machine.

Hardware

  • AMD Ryzen AI Max “Strix Halo”: Radeon 8060S-class iGPU, 12-core Zen 5/5c
  • 58 GB LPDDR5X shared with the GPU, 256 GB/s memory bandwidth (~3× the HX 370 tier); the GPU can address ~33 GiB
  • CachyOS

Inference stack

  • Lemonade 2026.44.1 serving llama.cpp b11393 on the Vulkan backend (Fedora 45 re-run pending — a newer llama.cpp build is expected to shift absolute times)
  • One LLM loaded at a time; models swap on demand
  • ROCm benchmarked vs Vulkan: no win for agent runs; Vulkan kept

Measured decode speeds (short bursts, same machine)

Model Generation MTP draft acceptance
Qwen3.6-35B-A3B-MTP 100 tok/s 87%
Qwen3.6-35B-A3B (no MTP) 61 tok/s —
Ornith-1.5-35B-A3B 106 tok/s 78%
Qwen3.8-27B dense 31 tok/s 84% (MTP)

Note: these are short-burst numbers on a 256 GB/s machine — they overstate long-batch throughput (see HX 370 power behavior for the mechanism; the Halo tier throttles less but the effect exists there too).

Results

The Strix Halo result table is being wired to the same auto-export pipeline as the HX 370 page. Until then, the seeded numbers live on Results — Strix Halo.