AMD has introduced the Instinct MI350P PCIe, a new enterprise AI accelerator aimed at companies that want to add AI capacity without replacing their current data center setup. The card is designed as a dual-slot, air-cooled PCIe part that fits standard 2U or larger servers and works within existing power, cooling, and rack infrastructure.

Built for Gradual AI Adoption

The MI350P is meant for organizations that want to start small and scale over time. Instead of deploying a full multi-GPU OAM system, customers can begin with a single card and expand as their AI needs grow.

That makes it a practical option for teams experimenting with on-premises inference, retrieval-augmented generation, and smaller AI models before committing to a larger hardware rollout.

Hardware and Performance

AMD says the MI350P includes 144GB of HBM3E memory with bandwidth of up to 4TB/s. It also supports lower-precision formats such as MXFP6 and MXFP4, along with sparsity acceleration, which can improve throughput by skipping zero values in matrices and data sets.

According to AMD, the card can deliver about 2,299 TFLOPS of performance, or up to 4,600 peak TFLOPS in MXFP4 mode. AMD says that makes it the fastest enterprise PCIe AI card currently available.

Workload Fit

AMD positions the MI350P for a wide range of inference workloads, including small, medium, and large AI models, plus RAG pipelines. The company says a single GPU can handle roughly 200 to 250 billion-parameter language models, while an eight-GPU node can support much larger inference deployments.

The card also works with AMD’s ROCm software stack, which is shared across its Instinct and Radeon product lines. That gives enterprises a familiar software path if they are already using AMD accelerators.

Why It Matters

The MI350P gives AMD a way to reach customers that are not ready for full rack-scale AI hardware. For many enterprises, the appeal will be lower friction: no major infrastructure redesign, no immediate eight-GPU commitment, and a clearer path to test AI workloads on existing servers.

AMD has not disclosed pricing or availability yet, so the practical impact will depend on how it is positioned once shipments begin.