Hey, got my 2x Arc B70 Pro setup working with vLLM 0.17.0-xpu. Still doing more testing and plan to do a full writeup this week with configs, docker-compose files, and detailed benchmarks, but here's what I've seen so far:

Hardware: 2x B70 Pro (32GB each), Ryzen 5 3600X, 48GB RAM, PCIe 4.0 x8, Ubuntu 24.04 w/ kernel 6.17

DeepSeek-R1-Distill-Qwen-32B (dense 32B, FP8 dynamic):

• \~22.5 tok/s single user (tensor parallel)
• \~14.6 tok/s single user (pipeline parallel)
• 16 GiB weights per GPU, 84K token KV cache
• TP is the clear winner for dense models on PCIe

Qwen3-30B-A3B (MoE 30B/3B active, FP8 dynamic):

• \~18-19 tok/s single user (both TP and PP similar, PP slightly faster)
• \~294 tok/s total throughput at 16 concurrent (PP) vs \~264 tok/s (TP)
• 14.5 GiB weights per GPU, 293K token KV cache
• PP actually wins for MoE on bandwidth-constrained PCIe — fewer inter-GPU transfers per token

Interesting finding: pipeline parallelism beats tensor parallelism for MoE models on PCIe 4.0 x8, but TP wins for dense models. Makes sense when you think about compute-to-communication ratio per layer. On NVLink TP would win both.

Getting the docker setup working was the hardest part honestly. Will document all the gotchas in the full post