A hands-on report of replacing Windows with Arch Linux ARM on an ASUS ProArt P16 H7607BA-SE033W (128 GB RAM, 6,144 CUDA cores, 2 TB SSD) to avoid Microsoft account requirements and the Windows-only official support. It documents collecting and applying NVIDIA kernel work, ACPI overrides, a patched hid-asus for keyboard support, and extracting ASUS speaker firmware, then packaging those changes into a bootable USB and an extended Archinstall. Kernel 7.0.0-1022-nvidia and NVIDIA driver 615.78.08 were used; display runs at 3840×2400@120Hz and the system uses NetworkManager, PipeWire, and rootless Podman with NVIDIA CDI. Build inputs are hash-pinned, hardware packages are signed with an Arch maintainer key, and the system reports a unified 116.0 GiB physical memory pool available to both the OS and CUDA.
Benchmarks show native ARM64 7-Zip performance roughly tied to Windows single-thread and about 5% faster compression and 3% faster decompression at 20 threads versus a comparable Windows machine. Linux reports about 6.3 GiB more usable unified memory for CUDA than a Windows setup with a large dedicated reservation. Running Qwen3.8-Flash-Next through vLLM with B12X CUDA kernels required ~90 GiB for model and KV cache, took ~16 minutes to load and compile, and delivered usable inference: code-review first token 0.316 s / complete 4.008 s, a 9,775-token incident log completed in 5.768 s, and a median generation rate ~30.9 tokens/s - comparable to Windows. Long-run stability testing (150 requests) showed minimal latency drift (<1.6%) and 100% correctness on deterministic checks.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.