Advanced Micro Devices (AMD) has announced the Instinct MI350P, a PCIe AI accelerator card designed for existing enterprise server racks [1], [4].
This launch allows companies to upgrade their AI inference capabilities without rebuilding their entire hardware infrastructure. By fitting into standard air-cooled servers, the card reduces the cost and complexity of deploying large-scale AI workloads [4], [6].
The MI350P is a dual-slot, drop-in upgrade that features 144GB of HBM3E memory [1], [2]. It is equipped with 128 compute units [3]. AMD said the card is roughly 40% faster in FP16 and FP8 theoretical compute compared to its competitor, the Nvidia H200 NVL [1].
Power requirements for the hardware are flexible to accommodate different data center environments. The card has a total board power of 600W [3]. However, it can be configured to run at 450W for deployments where power is constrained [3].
While the MI350P provides a streamlined path for enterprise adoption, it is a scaled-down version of other AMD offerings. The card contains half the cores, and half the memory, of the MI355X [2].
This design choice prioritizes compatibility with existing air-cooled systems over the raw performance found in more specialized, liquid-cooled AI clusters. The card is intended to make enterprise AI fit into the servers that companies already own [6].
“The card is roughly 40% faster in FP16 and FP8 theoretical compute compared to its competitor, the Nvidia H200 NVL.”
AMD is pivoting toward the 'brownfield' data center market by prioritizing physical and electrical compatibility over peak theoretical performance. By creating a card that fits existing air-cooled racks, AMD is lowering the barrier to entry for enterprises that cannot afford the massive capital expenditure required for the liquid-cooling systems often needed by the highest-end AI GPUs.
