AMD has officially disrupted the enterprise AI market with the launch of the MI350P, the first PCIe-based accelerator to feature 144GB of HBM3E memory. Built on the CDNA 4 architecture, the card aims to provide high-bandwidth inference capacity without the power and cooling overhead of OAM modules.
With ROCm 7.0 support and a peak performance of 4.6 PFLOPS in MXFP4, the MI350P is positioned as the primary competitor to Nvidia's H200 NVL for private AI clouds.