NorthstarAurora/strix-halo-bench-data

Strix Halo LLM Inference Benchmarks (gfx1151)

Reproducibility data for LLM inference benchmarks run on AMD Strix Halo hardware (Ryzen AI MAX+ 395 / Radeon 8060S iGPU / gfx1151, 128 GB unified memory). All logs are raw llama-bench and llama-server outputs from production benchmark runs documented in the strix-halo-llm-finetune-guide.

This dataset exists as a citation target — papers, blog posts, and forum threads referencing Strix Halo inference numbers can point to a stable Hugging Face URL instead of GitHub paths that may move.

What's in here

Directory	What it shows
rocwmma-fattn-sweep/	A/B benchmarks of GGML_HIP_ROCWMMA_FATTN=ON vs OFF on gfx1151. Finding: OFF wins by ~2.4× on prefill at 8k context for both dense Qwen3.5-27B Q8 and MoE Qwen3.6-35B-A3B Q4. Contradicts AMD's own RDNA 3.5 best-practices recommendation; matches Unsloth Studio's default.
vulkan-vs-rocm-sweep/	Head-to-head benchmarks of llama.cpp HIP backend vs Mesa RADV Vulkan backend at the same source commit (b9296). Finding: Vulkan wins Q4 decode by ~22%, ROCm wins BF16 decode by ~117%. Includes the vulkan-capability-extract.txt showing bf16: 0 on RADV STRIX_HALO that explains the inversion.
cublas-hipblaslt-sweep/	Build-flag sensitivity: with vs without GGML_HIP_GRAPHS, MMQ_MFMA, and the hipBLASLt vs CUBLAS-style path on b502 vs b867. Finding: the recommended flag combo (MMQ_MFMA=ON, NO_VMM=ON, GRAPHS=ON) is the right one and reproduces across two llama.cpp release tags.
qwen36-bench/	Qwen3.6-27B-MTP and Qwen3.6-35B-A3B raw inference benchmarks at multiple prefill/decode shapes. Used as the baseline for the MTP speculative-decoding write-up in the guide.

Hardware

• CPU: AMD Ryzen AI MAX+ 395 (16c/32t, Zen 5 + Zen 5c)
• GPU: Radeon 8060S iGPU (gfx1151 / RDNA 3.5), 40 RDNA CUs
• Memory: 128 GB unified LPDDR5x-8000 (BIOS UMA 1 GB, kernel auto-sizes GTT to full 128 GB)
• NPU: XDNA 2 (visible to rocminfo, not yet PyTorch-accessible)
• Platform: Corsair AI Workstation 300 (BIOS AXB35-02 3.07)

Software

• OS: Ubuntu 24.04 LTS
• Kernel: 6.19.14 mainline (KFD + amdgpu fence/dma_buf fixes for gfx1151)
• ROCm: 7.1.0 stable + nightly libhsa-runtime64.so.1 overlay from PyTorch's _rocm_sdk_core wheel (works around a null-pointer bug in stock 7.1.0 on gfx1151)
• Mesa: 25.2.8 (RADV STRIX_HALO support)
• llama.cpp: most runs at b9296 (commit a497476); cublas-hipblaslt sweep at b502 and b867 for build-flag comparison

Methodology

All runs use the same llama-bench command shape per sweep, with three repetitions and standard deviation reported. Mmap is explicitly disabled (--mmap 0) to remove cold-cache effects. Direct GPU dispatch is verified by checking ngl 99 reaches all layers.

For the ROCm path, every launch sets:

NIGHTLY_LIB=/srv/aurora-ai/venv/lib/python3.12/site-packages/_rocm_sdk_core/lib
LD_LIBRARY_PATH=$NIGHTLY_LIB:/opt/rocm-7.1.0/lib:/srv/aurora-ai/llama.cpp/build/bin

This is required to avoid the ROCm 7.1.0 HSA runtime null-pointer crash on gfx1151. Vulkan needs no overlay (Mesa is the driver).

For the build flags, ROCm builds use:

-DGGML_HIP=ON -DGGML_HIP_ROCWMMA_FATTN=OFF -DGGML_HIP_GRAPHS=ON \
-DGGML_HIP_MMQ_MFMA=ON -DGGML_HIP_NO_VMM=ON -DAMDGPU_TARGETS=gfx1151 \
-DCMAKE_HIP_FLAGS='--gcc-install-dir=/usr/lib/gcc/x86_64-linux-gnu/13'

Vulkan builds use:

-DGGML_VULKAN=ON -DGGML_NATIVE=ON

Full reproduction recipes, the rationale for each flag, and discussion of each finding live in the guide repo.

How to use

These are raw text logs intended for human reading, citation, and downstream parsing — not a structured dataset that datasets.load_dataset() will give you typed columns for. To consume programmatically:

from huggingface_hub import snapshot_download
path = snapshot_download(
    repo_id="NorthstarAurora/strix-halo-bench-data",
    repo_type="dataset"
)
# path/ now contains all the .log files

Each .log file is a verbatim capture of stdout+stderr from the bench command. The bench command itself is the first non-comment line of each log.

Citation

If you're referencing these numbers in a paper, post, or PR:

@misc{durkin2026strixhalobench,
  orcid        = {0009-0000-2537-1578},
  author       = {Paul Durkin},
  title        = {Strix Halo LLM Inference Benchmarks (gfx1151)},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/NorthstarAurora/strix-halo-bench-data}
}

Or just link the URL — informal references are fine.

Related

• ORCID: 0009-0000-2537-1578
• Guide repo: github.com/h34v3nzc0dex/strix-halo-llm-finetune-guide — full production guide; the source of truth that owns the methodology
• Long-form ROCm vs Vulkan article: articles/2026-05-rocm-vs-vulkan-strix-halo-precision-inversion.md
• Vulkan canonical dashboard: bench.ciru.ai — community-maintained Vulkan benchmark dashboard for Strix Halo and related AMD APUs

Contributing

Find an error or have a benchmark to add for a different model or different build flags? Open an issue or PR on the guide repo. New bench data gets folded back into this dataset on the next release.

License

MIT — see the LICENSE field. Use freely with attribution.