AMD Instinct MI350P - Why PCIe Data Center GPUs Still Matter

On May 7th, 2026, AMD released an addition to their data center GPU lineup: the AMD Instinct MI350P, a PCIe card! After seeing socketed data center GPUs from both leading GPU manufacturers, this is a welcomed edition to AMD Instinct’s lineup of GPUs and the first we’ve seen in nearly 5 years.

The Instinct MI350P brings AMD’s current generation GPU architecture to the data center via a PCIe based GPU. Why offer a PCIe data center GPU in the midst of OAM integrated GPUs? AMD is targeting mainstream server architectures for those who want to do on-premises AI inference, but either cannot high thermals and power density of OAM and SXM based nodes or cannot.

If you are looking for AMD Instinct MI350P data center GPUs, talk to our sales engineers today for more information! Configure an Exxact TensorEX server with AMD Instinct GPUs today.

AMD Instinct MI350P Specifications

AMD is essentially taking one of its MI350X accelerators and cutting it in half, resulting in a card with half as many compute units and half as much memory at just over half the power consumption.

Why Use a PCIe-Based Data Center GPU?

The majority of AI accelerator demand has centered on modular OAM and SXM-based GPUs designed for high-density compute nodes. These solutions deliver the best scale-up and scale-out performance through proprietary GPU interconnects. But they come with significant infrastructure requirements:

• 11kW per node
• Proprietary server design
• Low upgradability.

This has left a gap in the market for customers who need serious AI hardware without the overhead of high-density non-modular systems. GPU vendors have attempted to address this gap with workstation-class products repurposed for the data center. While capable, these graphics-based GPUs fall short in key areas:

• Lower peak performance compared to flagship server GPUs
• No HBM memory, limiting memory capacity and bandwidth
• Not a fully validated replacement for data center-class AI inference workloads

To properly fill this gap, AMD is releasing the Instinct MI350P — its first Instinct PCIe accelerator built on its current-generation CDNA 4 architecture since the MI210 in 2022. The goal is straightforward: deliver flagship-class AI hardware to traditional server and data center customers in a standard, readily replaceable PCIe form factor that can plug into existing server designs.

Instinct MI350P: Half a MI350X At Half the Power

The MI350P is effectively a precision half-slice of AMD’s flagship MI350X. Where the MI350X uses two I/O dies (IODs), each stacked with four accelerator complex dies (XCDs), for eight total XCDs, the MI350P uses a single IOD with four XCDs clocked identically to the MI350X. The result is roughly half the compute performance in a PCIe GPU.

The die reduction scales evenly across memory as well:

• Memory: 144GB HBM3e (half the XCDs mean half the 288GB on the MI350X)
• Memory Bandwidth: 4TB/s (half of the MI350X’s 8TB/s)
• Peak Matrix Performance (MXFP8): 2.3 PFLOPS vs. 4.6 PFLOPS

Physically, the MI350P is a standard full-height, full-length (FHFL) dual-slot PCIe card available in two TBP configurations: 600W standard and 450W optional. It is designed for air cooling and targets mainstream server form factors.

AMD did not include Infinity Fabric GPU-to-GPU interconnect support on the MI350P. Multi-GPU setups are limited to PCIe Gen5 x16 communication between cards, which caps usable AI model sizes to what fits within the 144GB of VRAM. This is the biggest difference from its MI350X OAM counterpart.

FAQ on AMD Instinct MI350P

What is the AMD Instinct MI350P?

The AMD Instinct MI350P is a data center GPU built in a standard PCIe add-in card form factor. It brings AMD’s CDNA 4 architecture, 144GB of HBM3E memory, and 2.3 PFLOPS of MXFP8 matrix performance to mainstream server platforms without requiring an OAM-based accelerator node.

What workloads is the AMD Instinct MI350P designed for?

The AMD Instinct MI350P is designed for enterprise AI inference, generative AI, retrieval-augmented generation (RAG), small and large language model deployment, and other accelerator-driven data center workloads. Its 144GB of HBM3E memory gives it enough capacity for many on-premises inference deployments while staying within a standard PCIe server design.

How much power does the AMD Instinct MI350P use?

The AMD Instinct MI350P has a 600W total board power configuration, with an optional 450W mode for systems with tighter power or thermal requirements. It is a passive, air-cooled full-height, full-length dual-slot PCIe card intended for compatible server platforms.

Who should consider using the AMD Instinct MI350P?

The AMD Instinct MI350P is a strong fit for organizations that want to deploy on-premises AI infrastructure but do not need, or cannot support, the density and power requirements of OAM or SXM GPU systems. It is especially relevant for teams looking to add high-capacity AI acceleration to existing PCIe-based server environments.

Conclusion

AMD has made significant strides in the data center GPU market, and with the Instinct MI350P, the company is finally returning to a long-underserved segment: the standard PCIe form factor. This is AMD's first Instinct PCIe accelerator built on its current-generation architecture in nearly five years, arriving at a time when demand for accessible, high-performance AI hardware is growing rapidly.

The MI350P is a purpose-built product, not a compromise. It won't match the raw throughput of the MI350X, but it delivers the highest performance AMD can package into a standard PCIe card — complete with CDNA 4 architecture, HBM3E memory, and data center-class capabilities — in a form factor that fits the infrastructure most organizations already have.