What if enterprises could scale AI workloads without rebuilding their data centers? A new PCIe accelerator approach enables inference deployments using existing infrastructure.
AMD is positioning its new Instinct MI350P PCIe cards as a way for enterprises to scale AI workloads using existing data center infrastructure, avoiding the costly power, cooling, and rack upgrades often associated with large GPU accelerator deployments.
The announcement targets a growing challenge facing organizations adopting AI at scale. While enterprises are moving AI initiatives from experimentation into production, many are finding that existing infrastructure is not prepared to support expanding AI workloads without significant new investment. Cloud-based AI deployments can also introduce concerns around data privacy, regulatory compliance, and unpredictable operating costs.
The AMD Instinct MI350P PCIe cards are designed to address those constraints by enabling AI inference deployments in standard air-cooled servers. The dual-slot accelerators can operate within existing power, cooling, and rack environments, allowing enterprises to expand on-prem AI capacity without redesigning their data centers or adopting dedicated accelerator platforms.
AMD says the cards are built for the emerging era of agentic AI and support workloads ranging from small AI models to large-scale inference deployments and retrieval-augmented generation (RAG) pipelines. Systems can support up to eight accelerator cards in standard air-cooled configurations, giving organizations a way to scale compute density while maintaining relatively simple deployment models.
The company is also emphasizing operational efficiency and flexibility. The MI350P PCIe cards include native support for lower-precision MXFP6 and MXFP4 formats designed for high-throughput AI inference, as well as sparsity acceleration support across mainstream 8-bit and 16-bit precisions.
AMD is additionally highlighting its open software ecosystem, including low- and no-cost development stack options intended to simplify deployment, reduce operational complexity, and lower overall AI infrastructure costs.
With the MI350P PCIe cards, AMD is targeting enterprises that want to expand AI capabilities on-prem while maximizing the value of existing infrastructure investments instead of undertaking large-scale data center upgrades.
Click here for the original announcement.
