The 128 GB Box

A Ryzen AI Max+ 395 with 128 GB of unified memory, set up headless as a local-LLM box.

A GEEKOM A9 Mega, installed before it ever booted Linux, then pushed until each layer showed what it actually does: a BIOS memory setting that takes RAM away from the GPU, an 85 W power preset that changes nothing, a benchmark bent by a download, a job queue that outlived two power cuts, and long-context prefill that took gpt-oss-120b 29 minutes to fill 128k, and hours on slower models. Then a blind code review and an honesty audit for the agents' runs. 30 posts, the numbers from the runs' logs.

  1. Installing Fedora on a computer that hadn't shipped yet
  2. 120 GiB for the GPU, set in the kernel, not the BIOS
  3. No BIOS image, so I diffed the NVRAM
  4. The download that contaminated the benchmark
  5. 128k context is the floor, and prefill is the wall
  6. Vulkan or ROCm on Strix Halo? Ask at which depth
  7. Mistral-Small-4 ran one matmul on the CPU, and said nothing
  8. The benchmark was talking to the previous server
  9. Read the transcript, not the gate
  10. IOMMU off buys a 128 GB iGPU 5 %, and costs the NPU
  11. The fastest model finished first by stopping first
  12. Inventing the work vs promising what isn't there
  13. The best-reviewed code came from the slowest Flash-Next arm
  14. Seven FAILs, one setting: rerunning agents at 262144 tokens
  15. 29 of 34 passing agent runs had a blocker the hidden tests never saw
  16. How to tell when your coding agent lied, not just failed
  17. Every way my GPU job queue went quiet
  18. A -> B -> C -> A: benchmarking a kernel cmdline change you have to reboot for
  19. One rebuild moved gpt-oss off ROCm. Two models stayed
  20. Empty diff, finished code: 124,000 characters of thinking
  21. The model said it saw no output. It was right.
  22. Three bugs between K2 and a working Spring app
  23. Three models wrote the same circuit-breaker bug
  24. The typo followed the tag; the failures didn't
  25. The lock map deleted a mutex someone was waiting on
  26. Nine accusations, one lie: re-reading round two's judges
  27. Five languages, two reps: the round-two ranking
  28. Two rounds, seventeen arms: the engine set the speed, not the quant
  29. The template differed. The prompts didn't.
  30. Eighteen days later, the same lie