HX 370 (Strix Point) / Fedora 45

The “realistic mini-PC” tier: what a developer can plausibly own and run agents on all day.

Hardware

  • AMD Ryzen AI 9 HX 370 (Strix Point): 12 cores (4× Zen 5 + 8× Zen 5c), Radeon 890M iGPU, XDNA2 NPU (58 TOPS, unused for LLM serving — see limitations)
  • 60 GB DDR5 dual-channel (~90 GB/s theoretical); the iGPU can address ~33 GiB of it
  • Mini-PC form factor, whitelabel AMI BIOS
  • Fedora 45

Inference stack

  • Lemonade 2026.45.3 serving llama.cpp b11846 on the Vulkan backend
  • One LLM loaded at a time; models swap on demand
  • ROCm backend installed but unused: not faster than Vulkan in agent runs
  • FLM/NPU backend measured separately: ~25% slower generation, ~2× slower prefill, continuation- only prompt cache — not viable for agent workloads

Per-model launch settings

Saved in Lemonade, quoted on each model’s page. Sampling follows each vendor’s model card. All models run at 64k context with --parallel 1.

Power behavior (important for reading our numbers)

The firmware runs AMD’s 65 W power skin: burst limits ~65 W, sustained limit ~54 W (STAPM). Measured draw during sustained decode sits exactly on that floor — the box is power-limit-bound, not thermally throttled (GPU 70 °C, CPU 73 °C, DIMMs ~76 °C vs a ~95–100 °C throttle point; DDR5’s 85 °C DIMM warning is the tightest margin).

Consequences, measured from long agent batches:

  • Sustained decode decays over minutes from fresh-clock speeds to a plateau of roughly 50–90% of fresh speed, depending on the model’s bytes-per-token (a 2.5B-active MoE held ~88%, a 4B- active MoE fell to ~50%). Recovery needs minutes of idle; short tool-call pauses don’t recover.
  • Prompt-reading speed decays ~40% under the same conditions.
  • Short raw-speed measurements (benchmarks of a few seconds) therefore overstate real agent-batch throughput ~2× on this machine.

The 54 W floor reads as the vendor’s deliberate 24/7-duty configuration. It is not a defect, and we benchmark it as-is.

Where the data comes from

Runs on this machine are exported from our private benchmark repo (llm-bench) as sanitized JSON and rendered at build time. See Results.