AMD unveils Instinct MI350P PCIe card

This is a practical move from AMD: instead of chasing only rack-scale deployments, it finally serves the large enterprise install base that needs serious GPU memory without redesigning the datacenter.

• –The 144GB HBM3E footprint makes the MI350P interesting for inference-heavy jobs that outgrow consumer cards but do not need OAM/SXM-style clusters
• –AMD’s choice to keep this as a PCIe deployment rather than a rack-scale OAM platform makes it more of a deployment-friendly inference card than giant model-sharding hardware
• –The 600W default and 450W fallback are the real story here: AMD is trying to fit current-gen accelerator performance into air-cooled enterprise servers
• –Open ROCm and the broader AMD software stack matter here because hardware alone will not move enterprise buyers
• –Lack of pricing and availability keeps it from being immediately actionable, but the spec sheet is strong enough to pressure Nvidia in PCIe datacenter deployments