... . --> Pascal Vulkan-vs-CUDA Cross-Machine Benchmark — 2026-05-06 — benchmarks.weeyuga.com
← All benchmarks
Headline Methodology Results Raw

6 MAY 2026 · pavilion+predator (cross-machine) · gemma/qwen3 · chat

Pascal Vulkan-vs-CUDA Cross-Machine Benchmark — 2026-05-06

26 calls across 26 cell(s)

Methodology

See HARNESS#pavilion-predator-vulkan-cuda-comparison-1 for the full procedure. Reproducible at git SHA ddbaaf46.

Results

Cell tok/s mean tok/s p50 tok/s p95 duration p50 calls
gemma4-e4b-q4km:cuda:ngl991
gemma4-e4b-q4km:cuda:ngl241
gemma4-e4b-q4km:cuda:ngl121
gemma4-e2b-q4km:cuda:ngl991
gemma4-e2b-q4km:vulkan:ngl991
qwen35-0.8b-q4km:cuda:ngl991
qwen35-0.8b-q4km:vulkan:ngl991
qwen35-2b-q4km:cuda:ngl991
qwen35-2b-q4km:vulkan:ngl991
qwen35-4b-q4km:cuda:ngl991
qwen35-4b-q4km:vulkan:ngl991
qwen35-0.8b-q4km:cuda:ngl991
qwen35-0.8b-q4km:vulkan:ngl991
qwen35-2b-q4km:cuda:ngl991
qwen35-2b-q4km:vulkan:ngl991
gemma4-e4b-q4km:cuda:ngl991
gemma4-e4b-q4km:vulkan:ngl991
gemma4-e2b-q4km:vulkan:ngl991
qwen35-4b-q4km:cuda:ngl991
qwen35-4b-q4km:vulkan:ngl991
qwen35-9b-q4km:cuda:ngl991
qwen35-9b-q4km:vulkan:ngl991
qwen3-14b-q4km:cuda:ngl991
qwen3-14b-q4km:vulkan:ngl991
qwen3-30b-a3b-iq2m-moe:cuda:ngl999:n-cpu-moe351
qwen3-30b-a3b-iq2m-moe:vulkan:ngl999:n-cpu-moe351

tokens per second — mean · p50 · p95

No tokens-per-second data captured.

Raw data

Every run gets its JSONL, log, summary, and metadata published. Clone the archive; re-run it; tell us where we got it wrong.

Cite

Margetic, S. et al. (2026). benchmarks.weeyuga.com/benchmarks/5111a3ee.html
Public benchmarks of the Weeyuga cluster. Run id: 5111a3ee-7d52-44d3-86cf-863b8d14e987. SHA ddbaaf46.

Related runs