I've been tuning llama.cpp on my Strix Halo box (Radeon 8060S, 128 GB) and the numbers are good enough to share. It's a branch called rdna35-perf on top of halo-box/strix-llama.cpp, a llama.cpp fork aimed at Strix Halo:

https://github.com/asf0/strix-llama.cpp/tree/rdna35-perf (10 commits, ROCm/HIP only)

TL;DR

• Prompt processing: 21–54% faster than Vulkan, and 8–174% faster than stock ROCm. The lead over ROCm grows with context, because stock ROCm flash attention falls off hard at depth.
• Generation: ties Vulkan on the dense 27B. On the MoE and the 2B it's 2–3% behind Vulkan at short context and within 1% from 64k up. 5–29% faster than stock ROCm.
• Quality: outputs stay closer to stock ROCm than Vulkan's do, and perplexity is unchanged.

Results

llama-bench, Q8_0 weights, q8_0 KV cache, flash attention on, tok/s. "Depth" is how many tokens are already in the KV cache when the test runs. Vulkan and ROCm are the same strix-llama.cpp master that rdna35-perf is built on.

Qwen3.6 35B-A3B (MoE, ub 4096)

Qwen3.8 27B (dense, ub 2048)

Qwen3.5 2B (dense, ub 2048)

f16 KV gives the same picture: PP 20–56% faster than Vulkan and up to +171% over stock ROCm at 128k, TG within −3% / +2% of Vulkan.

Quality

rdna35-perf changes kernels, so I checked it doesn't change outputs: KL divergence against stock ROCm on wikitext-2 (6 × 4096-token chunks), plus perplexity at 32k context.

rdna35-perf is closer to stock ROCm than Vulkan is, and perplexity is the same across all three builds.

What changed

• Flash attention kernel for RDNA3.5 (fattn-rdna35). This is the biggest win and the reason long-context PP holds up. GGML_FA_RDNA35=0 turns it off if you want to compare.
• Prefill GEMM tuning for the quantized matmuls.
• MoE: routing helper, fused reduction + residual, grouped mat-vec for speculative-decoding verify batches.
• Gated DeltaNet (the linear-attention layers in Qwen3.5/3.6/3.8): fused SiLU-gated output norm and fused decode.
• Decode: plain mat-vec for rows longer than 4096 (this closed the 27B decode gap to Vulkan), plus decode fusions for the Qwen3.6 MoE.
• Misc: HIP graphs off by default on RDNA3.5 (slower here), a faster penalties sampler, host-side prefill for Qwen3.5/3.6 with a draft model, and DFlash drafts read target features in place on the GPU.

I have only tried those 3 models so far.