Introduction

Work in progress: This wiki is being actively expanded from Discord research and fork release notes. Many tuning recipes and throughput numbers are community-reported and have not been independently reproduced here. See Verification status for what is solid vs what still needs checking on your hardware.

Welcome to the GFX1030 Wiki — a focused, hands-on knowledge base for running LLM inference on AMD gfx1030 (RDNA2 / Navi 21) GPUs, with a strong bias toward the Radeon PRO V620 and the purpose-built tooling collected here.

gfx1030 is the LLVM/ROCm target for the Navi 21 “Sienna Cichlid” die. It powers the consumer Radeon RX 6800 / 6800 XT / 6900 XT / 6950 XT and the workstation/data-center PRO W6800 and PRO V620. These cards are officially supported by ROCm on Linux, which makes them a cost-effective platform for modern LLMs — but getting the most out of them (power tuning, multi-GPU P2P, and RDNA-tuned kernels) takes a bit of extra work. That’s what this wiki documents.

Read it in this order

1. Getting started — hardware check, ROCm, optional V620 tuning, first container.
2. llama.cpp or vLLM — which stack matches the model and card count.
3. The matching section: vLLM onboard (V620) / vLLM overview, or llama.cpp.
4. Troubleshooting when a command fails. Verification status says which numbers are solid.

Host firmware (the machine never POSTs) and Flash-Next serve lines are specialist pages. Skip them until you need them.

What this wiki focuses on

• Tuning — the v620_toolbox recipes: lowering the V620’s VBIOS-locked 250 W floor to 120 W, enabling GPU↔GPU PCIe Peer-to-Peer between multiple V620s (Fedora + AMD hosts; power tuning also on Ubuntu 26.04; Intel Ice Lake often sees no P2P win — A/B it), and optionally disabling Pro-card ECC for ~2 GB extra VRAM.
• llama.cpp — stock builds plus the RDNA2 fork (tensor parallel, MTP / DFlash). Usually the easiest first stack on V620.
• vLLM on RDNA — ready-to-run Docker images; recipes for Hub vs recipe container, and a separate Flash-Next page (blivioniag/vllm-rdna on a blivioniag/rocm-rdna PyTorch base), how they are built with opengfx1030/vllm-rdna-docker, and the official vLLM extras fork opengfx1030/vllm-rdna (rdna_extras, now including Flash-Next cherry-picks). Hub v0.31.0-extras tracks the 11 Oct 2026 bake pin; commits after that pin still want a host venv clone.

New here? Start with Getting Started. Card identity is on Supported Hardware.

Community Discord

Much of this wiki is distilled from the gfx1030 club Discord — live tuning tips, fork releases, and multi-GPU benches land there first.

• Server: gfx1030 club
• Join: https://discord.gg/mESex2aBp
• Useful channels: #vllm-rdna, #llamacpp, #general, #benchmarks, #harnesses, #lmcache, #motherboard

When turning Discord threads into docs, follow the privacy rules in Contributing.

Disclaimer: This is a community wiki, not affiliated with or endorsed by AMD. Kernel patches and power-cap changes are done at your own risk. Always cross-check against the official ROCm documentation. Contributions welcome — see Contributing.