First-hand evaluation of AMD ROCm on Strix Halo reveals real-world viability of open-source alternatives to NVIDIA, a critical factor in diversifying the AI hardware ecosystem. Marco Inacio's primary account documents successful ROCm 7.2 installation on Ubuntu 24.04 requiring BIOS update for PyTorch GPU detection, reserved video memory set to 512 MB, GTT sharing for 128 GB unified pool, and GRUB_CMDLINE_LINUX_DEFAULT parameters ttm.pages_limit=32768000 plus amdgpu.gttsize=114688 (https://blog.marcoinacio.com/posts/my-first-impressions-rocm-strix-halo/). Setup enabled PyTorch 2.11+rocm7.2 via uv with custom index and llama.cpp server in Podman using HSA_OVERRIDE_GFX_VERSION=11.5.0 for Qwen3.6-35B-A3B inference with flash-attn.

Original coverage omitted quantitative latency or tokens-per-second data and long-term thermal stability under sustained load. AMD ROCm documentation details similar kernel and HSA overrides for new architectures while noting unified memory benefits for APUs (https://rocm.docs.amd.com/en/latest/). Phoronix reporting on prior consumer ROCm rollouts identified parallel early BIOS and driver friction patterns that delayed adoption by months (https://www.phoronix.com/news/AMD-ROCm-6.2-Release).

Strix Halo's CPU-GPU memory coherence, combined with llama.cpp and Opencode integration shown, connects to AMD's multi-year ROCm maturation since MI250 days and exposes NVIDIA CUDA's ecosystem lock-in as less absolute for local 35B-scale workloads. This configuration demonstrates fragmentation and addressing overhead trade-offs when mixing reserved and GTT memory are manageable, supplying developers a tested non-proprietary path that previous discrete-GPU ROCm coverage consistently under-reported.