vllm on intel-arc-pro-b70: Qwen3.5-27B fp8-e4m3 tg128 tok/s (bentech, dual B70+B60 vLLM FP8 pipeline-parallel)
13.3 tok/s (tg128, batch-1 per-request)
Fields
id
benchmarks/bentech-b70-qwen35-27b-fp8-dual-tgMetric
decode_tok_s
Value
13.3
Unit
tok/s (tg128, batch-1 per-request)
Model
Qwen3.5-27B
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. Qwen3.5-27B fp8-e4m3, dual B70+B60 vLLM FP8 pipeline-parallel.
Measured by
third_party
Reproducible
yes
notes
Qwen3.5-27B fp8-e4m3 on dual B70+B60 vLLM FP8 pipeline-parallel: pp512 1009.7 / tg128 13.3 tok/s. Dual GPU row-split over PCIe 4.0 x4 M.2 is SLOWER than single B70 for generation; value is context capacity not speed.
Notes
Qwen3.5-27B fp8-e4m3 on dual B70+B60 vLLM FP8 pipeline-parallel: pp512 1009.7 / tg128 13.3 tok/s. Dual GPU row-split over PCIe 4.0 x4 M.2 is SLOWER than single B70 for generation; value is context capacity not speed.