citable URL: https://halobench.com/records/clm-0070/ — this address never moves; the anchor /records/#clm-0070 keeps resolving
On gfx1151, Ornith-1.0-35B Q8_0 is measured under the per-model, per-phase backend rule (clm-0050), and Vulkan wins outright here -- unlike laguna-s-21 (ROCm-favoured, clm-0065) but matching nemotron3-super and deepseek-v4-flash's Vulkan-favoured pattern: Vulkan leads decode at every measured depth (55.70 vs 47.66 t/s at d0, +17%; 46.22 vs 39.75 at d32768, +16%; 32.37 vs 27.48 at d131072, +18%) and prefill at d0/ d32768 (1063.7 vs 865.3, +23%; 679.2 vs 525.2, +29%), with ROCm only marginally ahead on prefill at the deepest cell (244.8 vs 238.5, +2.6%). Most notably, Vulkan did NOT lose the GPU device at d131072 on this candidate -- a clean break from the laguna-s-21/deepseek-v4-flash device-loss precedent (clm-0054/clm-0059/clm-0065) that has held on every other hybrid/MoE model tested at this depth on this box to date. Vulkan is the backend this bench serves the tau2 capability arm on.
METHOD -- ornith-35b-fullbench matrix: stock llama.cpp 3653e6d (ROCm) / 3653e6d6d (Vulkan, prefix-matches per house convention), default f16 KV (matches the 2026-08-16 screen's serving config, cfg-0093), fa on, -ngl 999, --load-mode none, pp1024/tg256, llama-bench defaults otherwise (-b 2048/-ub 512, not passed explicitly). Output sanity gate (chat-endpoint coherence check) passed for both backends before any throughput was recorded. N=3 fresh-process reps at d0/d32768 (max CoV 0.9% rocm-d0-pp, 0.2% vulkan-d32768-pp -- well inside the 3% scatter-flag threshold, a healthy code path per protocol §0), N=1 at d131072 per protocol Tier1 / deepseek-v4-flash convention.
pp1024 / tg256 by depth (mean of N):
| depth | ROCm | Vulkan | vk/rocm pp | vk/rocm tg | |---|---|---|---|---| | 0 | 865.28 / 47.66 | 1063.66 / 55.70 | 1.229 | 1.169 | | 32,768 | 525.19 / 39.75 | 679.22 / 46.22 | 1.293 | 1.163 | | 131,072 | 244.84 / 27.48 | 238.54 / 32.37 | 0.974 | 1.178 |
Vulkan's d131072 cell (run-0343/run-0344, N=1) needed only its first attempt out of the 2-attempt device-loss cap and produced a clean, coherent JSON result -- no vk::DeviceLostError, no GPU device wedge, no dmesg evidence of a ring reset in the window (checked via `sudo dmesg -T`, per the house rule fixing the non-sudo capture bug noted on clm-0065: this job's own on-failure dmesg calls are sudo-prefixed, though they were never triggered since no failure occurred). Whether this holds at contexts deeper than 131,072 is untested; d131072 is this bench's own matrix ceiling.
No MTP/speculation on either backend: this job's own from-scratch GGUF header parser (no numpy/gguf-py dependency available on-box, so a minimal pure-Python GGUF v3 reader was written) read all 40 blocks / 733 tensor names directly and found zero nextn/mtp/ eagle/medusa/draft tensors -- plain decode throughout, matching the tau2 arm.