llama-cpp on intel-arc-pro-b70: Gemma 4 31B gguf-q4-k-m tg128 tok/s (bentech, single B70 SYCL)
18.6 tok/s (tg128, batch-1 per-request)
Fields
id
benchmarks/bentech-b70-gemma4-31b-q4km-sycl-tgMetric
decode_tok_s
Value
18.6
Unit
tok/s (tg128, batch-1 per-request)
Model
Gemma 4 31B
Methodology
THIRD-PARTY MEASURED (Ben Linville/bentech, Apr 16 2026, llama-benchy 0.3.5, 25-run matrix). Arc Pro B70 32GB (BMG-G31) + B60 24GB (BMG-G21); Ryzen 7 7800X3D, 64GB DDR5, B850 PCIe4 x4 M.2 adapters; llama.cpp b1-ff5ef82 Vulkan+SYCL, Intel llm-scaler-vllm:0.14.0-b8.1 Docker (vLLM FP8 pipeline-parallel). pp512/tg128, depth0, 3 runs/test mean, no prefix caching, q8_0 KV, FA on, 8k ctx. Per-request tg, NOT aggregate. Gemma 4 31B gguf-q4-k-m, single B70 SYCL.
Measured by
third_party
Reproducible
yes
notes
Gemma 4 31B gguf-q4-k-m on single B70 SYCL: pp512 204.7 / tg128 18.6 tok/s. Single card. Q4_K_M fastest generation for the quant sweep on this hardware.
Notes
Gemma 4 31B gguf-q4-k-m on single B70 SYCL: pp512 204.7 / tg128 18.6 tok/s. Single card. Q4_K_M fastest generation for the quant sweep on this hardware.