Every number on this page comes from one machine, a Strix Halo box with 128GB unified memory, under a recorded configuration, with the build hash and the requested flash-attention mode logged from the benchmark's own JSON. Rows are added when a build is worth measuring, not on a fixed schedule. For breadth across many stacks, the community toolbox benchmarks cover more ground; this page trades breadth for controlled depth: pinned builds, same-base comparisons, a context-depth axis, and drift stated instead of smoothed.

Drift note: Deep dense prompt-processing cells vary a few percent day to day on this hardware. Deltas within a run matter more than absolute digits across days.

Configuration legend: cfg-A means amd_iommu=on iommu=pt, used before 2026-07-18. cfg-B means amd_iommu=off, the current production configuration. cfg-C means Fedora 44, kernel 7.1.5, Mesa/RADV 26.1.5, amd_iommu=off, Vulkan only, and production build 10215 with nine curated patches.

Config epoch C opened on 2026-07-31. The reference box moved to Fedora 44, kernel 7.1.5, and Mesa/RADV 26.1.5; ROCm was removed, and production moved to the nine-patch Vulkan build. ROCm rows are historical because the box no longer carries a ROCm stack.

Reference Set v1 is unchanged. cfg-B and cfg-C rows use the same two model files, so they are comparable model-wise even though the platform and production build differ.

Reference Set v1 (July 2026)

Qwen3.6-35B-A3B UD-Q5_K_XL (MoE)

Qwen3.6-27B UD-Q5_K_XL (dense)

The set changes only by explicit version bump. Every bump gets a changelog entry, a full re-baseline, and the old set's table archived on this page.

cfg-C reference results

All cfg-C rows use f16 K/V, batch 2048, ubatch 512, mmap off, prompt length 512, generation length 128, and depths 0 and 32768. Production commit 6d78899ea, build 10215, is base b10206 plus nine curated community patches. The patched provenance is part of every row set and must not be read as stock upstream behavior.

Qwen3.6-35B-A3B UD-Q5_K_XL

Date Build Commit Backend Model and quant FA requested Config Depth pp512 t/s tg128 t/s
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-C 0 1194.116 56.721
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-C 32768 761.794 47.042
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 0 cfg-C 0 1172.101 56.680
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 0 cfg-C 32768 510.764 41.787

Qwen3.6-27B UD-Q5_K_XL

Date Build Commit Backend Model and quant FA requested Config Depth pp512 t/s tg128 t/s
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C 0 353.908 10.345
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C 32768 251.952 9.308
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-27B UD-Q5_K_XL 0 cfg-C 0 344.766 10.327
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-27B UD-Q5_K_XL 0 cfg-C 32768 182.743 7.983

What the prefill patch changes

These same-base attribution builds use Qwen3.6-27B UD-Q5_K_XL, f16 K/V, Vulkan, and flash attention on under cfg-C. The community-patched variants are under upstream discussion and do not describe stock upstream behavior.

Date Build Commit Backend Model and quant FA requested Config Variant pp@32k t/s pp@65k t/s
2026-07-31 10206 69e62fc77 Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C Stock upstream b10206 94.457 29.803
2026-07-31 10214 47f93a822 Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C Contiguize removed 134.537 45.095
2026-07-31 10214 1e355da72 Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C Hoist removed 242.170 180.738
2026-07-31 10212 7ad21320f Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C Stock plus contiguize and prerequisites 237.354 180.390
2026-07-31 10215 6d78899ea Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-C Full nine-patch production 251.979 198.162

The owning change is contiguize strided f16 KV for FA prefill: removing it brings the dense cliff back, while adding it with its prerequisites removes the collapse.

Generation is unaffected, staying within 0.09% across all five binaries.

On the collapsing path, repeated deep dense cells scatter by 5 to 10%. On the fixed path, the same cells reproduce within about 0.3%. Repeat deep dense runs before comparing small deltas.

Archive: earlier configurations

All rows use batch 2048, ubatch 512, mmap off, prompt length 512, generation length 128, and depths 0 and 32768.

Qwen3.6-35B-A3B UD-Q5_K_XL

Date Build Commit Backend Model and quant FA requested Config Depth pp512 t/s tg128 t/s
2026-07-23 9870 2d973636e ROCm Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-B 0 1089.106 49.313
2026-07-23 9870 2d973636e ROCm Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-B 32768 608.617 42.831
2026-07-18 9870 2d973636e Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 0 cfg-B 0 1046.945 56.597
2026-07-18 9870 2d973636e Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 0 cfg-B 32768 483.272 41.578
2026-07-18 9870 2d973636e Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-B 0 1076.145 56.780
2026-07-18 9870 2d973636e Vulkan Qwen3.6-35B-A3B UD-Q5_K_XL 1 cfg-B 32768 685.943 46.992

Qwen3.6-27B UD-Q5_K_XL

Date Build Commit Backend Model and quant FA requested Config Depth pp512 t/s tg128 t/s
2026-07-23 9870 2d973636e ROCm Qwen3.6-27B UD-Q5_K_XL 1 cfg-B 0 375.528 9.977
2026-07-23 9870 2d973636e ROCm Qwen3.6-27B UD-Q5_K_XL 1 cfg-B 32768 223.078 9.101
2026-07-18 9870 2d973636e Vulkan Qwen3.6-27B UD-Q5_K_XL 0 cfg-B 0 342.351 10.362
2026-07-18 9870 2d973636e Vulkan Qwen3.6-27B UD-Q5_K_XL 0 cfg-B 32768 181.652 8.001
2026-07-18 9870 2d973636e Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-B 0 350.971 10.381
2026-07-18 9870 2d973636e Vulkan Qwen3.6-27B UD-Q5_K_XL 1 cfg-B 32768 97.042 9.336

Changelog

2026-08-05: production moved to build 10283 (b7b85da9), the full strix-halo-vulkan branch tip. The cfg-C tables below stay as measured on build 10215; a cfg-D reference set follows once it is re-baselined.
2026-07-31: Added cfg-C (Fedora 44, Mesa 26.1.5, production build 10215 with nine curated patches); earlier cfg-A and cfg-B tables moved to the archive section; added the dense prefill patch attribution table.
2026-07-23: page created, initial data from tuning series parts 1-2 and the KV verification