4x RTX 3090 vLLM Qwen3.5-27B 50-request aggregate (Level1Techs reference run)

THIRD-PARTY MEASURED (wendell / Level1Techs). vLLM, 4x 3090 tensor-parallel, 50 concurrent requests: ~348 t/s avg output. The same test on 4x Arc Pro B70 TP=4 returned 370 t/s avg / 550 peak (comm-arc-pro-b70-vllm-qwen35-27b-decode) - the B70 leads on batched aggregate, the 3090 leads on per-request decode.

Reference baseline for the B70 multi-card number. AGGREGATE throughput, not per-request: per-user decode is far lower at 50 concurrent. ~$3,600 of used cards + host.