One AMD driver flag boosts dual-GPU Vulkan LLM inference up to 4x
tabletuser_blogspot · reddit · 2026-09-28
A Reddit user running llama.cpp with the Vulkan backend on dual Radeon GPUs (MI50 16GB + RX 7900 GRE) diagnosed why benchmarks were low and found a major fix:
- Hardware layout: moving the MI50 to the primary PCIe 16x slot and demoting the RX 7900 GRE to a 4x slot improved overall inference performance.
- Key flag: RADVPERFTEST=nogttspill — an AMD Linux driver flag that forces Mesa's RADV Vulkan driver to keep model allocations in VRAM instead of spilling into system memory (GTT).
- Results: pp512 gains up to +402% (Laguna-XS-2.1), +398% for gemma-4-31B, +387% for Muse-Glimmer-30B (which also gained +48.6% tg128); a few MoE models like Qwen3-Coder-30B-A3B saw marginal losses.
- The author used Google Gemini to help tune llama-bench settings and shared full commands plus a 13-model benchmark table.
Anyone running local LLMs on AMD GPUs should try this flag immediately.
More from Infra
- India's Sarvam launches Token Factory, hosting top open models on local infra — HarveenChadha · 2026-11-16
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16