Intel LLM-Scaler 1.2 on 4x Arc Pro B60 (TP=4): Qwen3-VL-30B decision-support 1K->2K TPOT (c=16)

THIRD-PARTY MEASURED (Embedded LLM, Feb 14 2026). Sparkle 4x Intel Arc Pro B60 workstation, TP=4, 24GB/456GB-s per GPU. Qwen/Qwen3-VL-30B-A3B-Instruct. vLLM 0.12 vs Intel LLM-Scaler 1.2. Concurrency 16/32/64, 1K->1K and 1K->2K scenarios. Decision-support 1K->2K at concurrency 16, Intel-optimized LLM-Scaler (vLLM+Intel opt).

LLM-Scaler (Intel opt vLLM) TPOT 44ms vs stock vLLM 59ms on same hardware - 25% decode-stream gain from Intel kernels.