Supported Hardware

gfx1030 is the LLVM target name for AMD’s Navi 21 GPU (RDNA2 generation, codename Sienna Cichlid). Several retail and workstation cards use this die and therefore report themselves as gfx1030.

Cards that natively report as gfx1030

Pro cards hide ~2 GB behind ECC. See Disabling ECC if you want the full 32 GB.

#general (Sep 15 2026): versus a 32 GB MI50 (gfx906), community first-pass on V620 was more prefill, less decode — PP “almost 2×”, TG behind but not far. That is a community A/B, not a wiki bench. Sibling notes: wiki-gfx906.

#general (Sep 17 2026): do not tensor-parallel a V620 with an NVIDIA card (community example: RTX 5090 + V620). Opposite strengths (RDNA2 has no matrix cores; the NVIDIA card has little VRAM per dollar) — mixed decode gains were small and prefill collapsed. Prefer a homogeneous V620 (or W6800) box, or run the NVIDIA card standalone.

Do not flash W6800 VBIOS onto a V620. #general (Sep 2026): pre-modded eBay cards and DIY flashes that load a W6800 BIOS on a V620 drop active CUs from 72 → 54 (W6800’s different CU layout). That is a large compute hit — stay on stock V620 BIOS. Cooling shroud mods are fine; BIOS swaps are not a free “Pro unlock.”

Need OverDrive / unlocked clocks in a passthrough VM without flashing? Prefer the reversible 4-byte PowerPlay soft unlock (Tamalero/amd-v620-soft-unlock) over a W6800 flash — see Power tuning → Soft unlock.

All of these are on the officially supported list for recent ROCm releases on Linux.

AMD BC-250 (gfx1013)

Headless mining APU board (Cyan Skillfish). RDNA2, reports as gfx1013 — not Navi 21 / gfx1030. Community ROCm and Vulkan work exists; most of this wiki still targets Navi 21.

Community references: AMD BC-250 docs, akandr/bc250-rocm.

RDNA2 relatives that can run gfx1030 code

The rest of the RDNA2 line uses a different LLVM target but shares the same ISA family. They are not on the official support matrix, but in practice they run gfx1030 kernels once you set HSA_OVERRIDE_GFX_VERSION=10.3.0 (see HSA_OVERRIDE for RDNA2 Cousins).

Architecture highlights (Navi 21)

• RDNA2 compute units with a native wavefront size of 32 (wave32), unlike GCN’s wave64.
• Infinity Cache (128 MB on Navi 21) that dramatically raises effective memory bandwidth.
• No dedicated matrix/tensor cores — RDNA2 predates the WMMA/matrix instructions added in RDNA3 (gfx11xx). Matrix math runs on the regular vector ALUs, so expect lower peak throughput than RDNA3 or CDNA cards, but very good price/performance for inference.
• FP16 is well supported; BF16 has limited/emulated support and is best avoided for hot paths (prefer float16). #general (Sep 28 2026): ComfyUI / image-video workflows often default to bf16. On V620 that path is HIP-emulated, so those pipelines are a poor fit — pick fp16 / int8 workflows, or use a card with native bf16. Community also called video generation a CUDA-world problem, not a V620 win. #general (30 Oct): training on V620 is the same constraint — prefer FP32 (or a native-bf16 card); community bf16-emulation tests had drawbacks.

How to check your GPU target

rocminfo | grep -i 'gfx\|Name'
# or, more directly:
rocminfo | grep -m1 -o 'gfx[0-9]*'

If the output shows gfx1030, everything in this wiki applies directly. If it shows gfx1031, gfx1032, etc., head to HSA_OVERRIDE for RDNA2 Cousins first.

Host firmware (large BAR / no POST)

4× V620 behind a PLX switch can fail to POST on older workstation BIOS (community: Dell Precision T5820 / T7820 / T7920) even when the same kit boots on a modern desktop board. That is firmware MMIO High / SR-IOV work — see Host firmware — not a ROCm bug.