Skip to content

Hardware overview

hal0 runs on four classes of home hardware in v1. The one-line installer detects which class you’re on and picks the right backend automatically — this page is for figuring out where you land before you commit.

TierHardwareStatusPath
First-classAMD Strix Halo (Ryzen AI Max+ 395, Radeon 8060S, XDNA NPU, 64–128 GB unified)Reference platformVulkan llama.cpp on the iGPU + FLM-on-NPU (NPU pending)
SupportedAMD discrete (RX 7900 XT/XTX, Radeon Pro)Vulkan today, ROCm queuedVulkan llama.cpp; ROCm toolbox pending
SupportedNVIDIA discrete (RTX 3080 / 4080 / 4090 / 5090)Vulkan today, CUDA queuedVulkan llama.cpp; CUDA toolbox pending
FallbackCPU-only x86_64 (no GPU)CI smoke targetVulkan-CPU (lavapipe)

Linux + systemd is required for all tiers.

  • You’re shopping for a home AI box. Get a 128 GB Strix Halo machine. That’s the whole premise of v1 — unified memory means you run the big models that discrete cards can’t, in a single quiet SFF chassis.
  • You already have a high-end NVIDIA card. Use it. The NVIDIA page covers what works today (Vulkan) and what’s queued (CUDA toolbox). A 4090 / 5090 on chat tok/s outperforms an iGPU on small models; you trade headroom for throughput.
  • You already have an AMD discrete card. Same story. AMD discrete — Vulkan today, ROCm toolbox queued.
  • You have a CPU-only box and want to try hal0. CPU-only walks through the fallback path. Smoke-test it; don’t expect to chat through it all day.

Every measured number on this site comes from the Strix Halo reference deployment (Ryzen AI Max+ 395 + Vulkan llama.cpp). See the Strix Halo page for the verbatim table. Numbers from other hardware tiers will be added as they’re measured at publish quality.