Chip designer AMD unveiled its latest line of Peripheral Component Interconnect Express (PCIe) cards to help AI infrastructure operators boost compute performance.

The Instinct MI350Ps are dual-slot drop-in cards for standard air-cooled servers, with users able to slot the hardware into their existing stack.

Available in air-cooled systems with up to eight accelerator cards, the chip giant claims its cards can support some of the highest performance loads for an enterprise PCIe card, with up to 4,600 peak teraflops at micro-scaling four-bit floating point (MXFP4).

AMD touts the device as yet another extension of its already stacked AI-centric product lineup, claiming the performance increases from the MI350P make it “ideal for small, medium, and large AI models for inference and retrieval-augmented generation (RAG) pipelines.”

“MI350P PCIe cards support the spectrum of precision levels that enterprise AI models rely on most,” Suresh Andani, AMD’s corporate VP for compute and enterprise AI, wrote in a blog post. “Adopting AI doesn’t mean rebuilding infrastructure from the ground up. With MI350P cards, enterprises can run more models and serve more users within their existing data centers.”

PCIe cards are high-speed interfaces that connect compute components, like graphic processing units (GPUs) or accelerators, directly to a server's motherboard or central processing unit (CPU). It acts as a fast lane built into the server itself, allowing add-in cards to slot into existing machines without requiring custom infrastructure. In the case of complex AI agentic workloads, that direct connection to the CPU is a real boon.

AMD’s mammoth Xilinx acquisition saw it take control of an established PCIe accelerator lineup, which it has since sought to push to the high-performance computing (HPC) and later enterprise AI markets.

The firm’s software stack then comes in to provide users with the means to migrate workloads with minimal code changes while also extending support for life cycle tools like the Kubernetes GPU Operator.

“When combined with Instinct MI350P cards and partner-delivered solutions, the stack enables organizations to get up and running quickly on-premises without ongoing per-token charges,” Andani added.

AMD’s PCIe cards augment its existing infrastructure peripheral lineup that was boosted by its $1.9 billion acquisition of data processing unit (DPU) vendor Pensando in 2022. AMD has since offered device designs addressing growing bottlenecks in hyperscale environments, with workloads becoming increasingly intense in both complexity and scale.

Among the customers lining up for its latest AMD addition are Lenovo, Cisco, Dell Technologies, and Hewlett Packard Enterprise (HPE).

“Our work with AMD on Instinct MI350P reflects the growing need to tightly integrate AI compute with intelligent networking and security," Cisco compute GM and SVP Jeremy Foster noted in a statement. "MI350P creates new opportunities to deliver scalable, resilient AI infrastructures that connect data, models, and users across the enterprise."

“The MI350P's footprint and memory profile make it a compelling option for distributed inference workloads, where bringing intelligence closer to users is essential,” Shawn Michels, VP of product for cloud computing at Akamai, added. “We're encouraged by AMD's continued innovation in accelerator design, and the role purpose-built inference silicon will play in enabling agentic AI at the edge.”