Strix Halo llama.cpp branch boosts prompt processing 21–54% over Vulkan
asfbrz96 · reddit · 2026-10-12
A developer tuning llama.cpp on Strix Halo (Radeon 8060S, 128 GB) released the rdna35-perf branch (10 commits, ROCm/HIP only) on top of halo-box/strix-llama.cpp:
- Prompt processing: 21–54% faster than Vulkan and 8–174% faster than stock ROCm, with the ROCm lead growing with context depth (stock flash attention falls off hard); at 128k the 27B dense model is +174% over stock ROCm
- Generation: ties Vulkan on the 27B dense model; 2–3% behind on MoE/2B at short context, within 1% from 64k; 5–29% faster than stock ROCm
- Quality: KL divergence vs stock ROCm lower than Vulkan's, perplexity unchanged
Tested with Qwen3.6 35B-A3B (MoE), Qwen3.8 27B, and Qwen3.5 2B on Q80 weights with q80 KV cache, full benchmark tables included.
More from Infra
- Micro TECs for ELS optical modules could be a $200-600M market by 2028 — jwt0625 · 2026-10-12
- Yandex bot traffic drops 99% after data center attack, Cloudflare Radar shows — rooftopzen · 2026-10-12
- Bit-exact speculative decoding hits 106 t/s on Qwen3.8-27B with one RTX PRO 6000 — stargate425 · 2026-10-12
- Chinese chipmaker CXMT hits design breakthrough in tech sought by Samsung, SK Hynix — AIFlow_ML · 2026-10-12
- China's Infrawaves raises $322M Series C, hitting 43B yuan total — record for domestic AI infra — 智东西 · 2026-10-12
- Same model, different behavior: Reddit asks how to test open-model inference compatibility — Top_Speaker_7785 · 2026-10-12