Surprisingly Useful AI Article Enhancements
AMD has introduced the Instinct MI350P PCIe card, a dual-slot GPU inference accelerator designed for enterprises that want to run AI workloads on premises without overhauling their data centre power, cooling or rack infrastructure.
The card is positioned as a middle path between cloud-based AI – which can raise privacy concerns and introduce unpredictable costs – and large-scale GPU accelerator platforms that often require expensive data centre redesigns.
Fitting AI into existing infrastructure
The MI350P is built as a drop-in card for standard air-cooled servers. AMD is targeting organisations that need more AI computing power than CPUs alone can deliver but are not ready to commit to dedicated GPU accelerator platforms.
The cards can be deployed in air-cooled systems with up to eight accelerators per server, making them suited to small, medium and large AI models for inference and retrieval-augmented generation (RAG) pipelines.
“As organisations adopt AI, many discover that their infrastructure struggles to keep up,” AMD stated in its product blog. “Running AI in the cloud is an option, but the cloud can introduce privacy concerns and unpredictable costs. Upgrading on-prem infrastructure is another option, but supporting large GPU-accelerator platforms can require expensive redesigns in data centre power and cooling.”
The MI350P is intended to offer enterprises a way around that trade-off, according to AMD, by delivering GPU-accelerated inference within the constraints of standard rack environments.
Performance specifications
AMD has outlined several performance claims for the MI350P. The card provides native support for lower-precision MXFP6 and MXFP4 data formats, which are designed to deliver high throughput for inference workloads. It also supports sparsity acceleration for most mainstream 8-bit and 16-bit precisions.
AMD estimates the card can reach 2,299 teraflops (TFLOPS) and up to 4,600 peak TFLOPS at MXFP4, which the company describes as the highest performance currently available in an enterprise PCIe card form factor.
The MI350P is equipped with an estimated 144GB of high bandwidth memory 3e (HBM3E) running at up to 4TB/s.
Support for FP8, MXFP8 and MXFP4 precision formats is a key reason the card can process current AI workloads within standard air-cooled data centres, according to AMD. Higher precision formats such as INT8 and BF16 benefit from the card’s sparsity support to deliver efficient performance while reducing memory usage and power demands.
Open software ecosystem
AMD is positioning the MI350P within its broader open AI software strategy. The card is supported by the open-source ROCm ecosystem and what AMD calls its enterprise AI reference stack, which is provided to partners at no licensing cost.
The software stack includes the Kubernetes GPU Operator for lifecycle management, cloud-native AMD Inference Microservices and native support for AI frameworks such as PyTorch. AMD argues this enables organisations to migrate inference workloads with minimal code changes.
“We built AMD Instinct MI350P PCIe cards with open standards for cross-platform interoperability,” the company outlined in its blog. “Their addition continues our strategy of enabling a fully open AI ecosystem and providing customer choice in enterprise environments.”
The absence of ongoing per-token charges is another point AMD is emphasising. When combined with the MI350P hardware and partner-delivered solutions, the open-source stack is designed to let organisations run AI on premises without the recurring costs associated with cloud-based inference services.
Scaling from pilot to production
AMD is framing the MI350P as part of a broader enterprise AI portfolio that gives organisations a path from AI evaluation through to production deployment without vendor lock-in or architectural compromise.
Related: Best Business Laptops for work & school
Related: Best Gaming Laptops
Related: Best Portable Laptop
The company argues that adopting AI should not require rebuilding infrastructure from the ground up. The MI350P is designed to allow enterprises to migrate workloads without code rewrites, integrate with existing AI pipelines and scale as workloads evolve.
For organisations currently running AI pilots or proof-of-concept projects, the PCIe form factor offers a lower barrier to entry than full-scale GPU accelerator platforms. The drop-in design means enterprises can begin deploying inference capabilities using their current server infrastructure and expand as demand grows.
AMD has not disclosed Australian pricing or specific availability dates for the Instinct MI350P PCIe card.
Last Updated on May 10, 2026 by Nick Ross
