Intel Arrow Lake CPUs cripple multi-GPU AI setups with broken P2P

A tech enthusiast building a multi-GPU machine for AI inference tested an Intel Core Ultra 7 270K (Arrow Lake) on an Asus Z890 Apex motherboard with two NVIDIA RTX A6000s. While the CPU provides 24 PCIe 5.0 lanes and supports 8x8x bifurcation for dual x16 slots, the actual P2P performance was abysmal. Unidirectional P2P bandwidth dropped from 619.30 GB/s (disabled) to just 5.48 GB/s with P2P enabled. Bidirectional bandwidth fell from 16.60 GB/s to 10.96 GB/s. These numbers came only after installing a patched open-gpu-kernel-modules driver, which itself points to a fundamental issue.

Further investigation revealed a hardware/firmware limitation in Arrow Lake's PCIe root complex that prevents correct packet routing between GPUs (GitHub issue #1253). The problem is not driver-level; it's baked into the chipset. The author notes that even if P2P connectivity registers as 'possible', the actual bandwidth is halved, making multi-GPU setups on Intel consumer platforms largely useless for training or inference tasks that rely on fast GPU-to-GPU transfers. For AI builders, the recommendation is clear: use AMD Threadripper, Intel Xeon/EPYC, or employ a PCIe switch board to bypass the root complex. This finding has immediate implications for the growing DIY AI hardware community.