2 min read
Add as a preferred source on Google

AMD Adds Instinct MI350P Support to Windows Server for AI Inference

The support covers Windows Server 2025 and the upcoming Windows Server 2028. GPU partitioning can divide an accelerator’s capacity among workloads.

A GPU card rests beside an open server chassis. / TokenPost.ai
A GPU card rests beside an open server chassis. / TokenPost.ai

AMD announced support for Instinct MI350P GPUs on Windows Server 2025 and the upcoming Windows Server 2028, giving enterprises a way to run AI inference in their own data centers.

The support covers Windows Server 2025 and the upcoming Windows Server 2028.

AMD Inference Microservices can simplify deployment of supported models. GPU partitioning would let IT teams divide an MI350P’s capacity among workloads rather than assign the entire accelerator to one model.

Running inference on premises would give companies control over where sensitive information, such as proprietary code and customer records, is processed. Shared inference services could support multiple internal applications while allowing teams to allocate GPU capacity to each workload.

The plan addresses a constraint for data centers adding AI: new systems must fit alongside existing applications within limits on budgets, rack space, power and cooling. AMD positions its EPYC processors as a way to consolidate traditional workloads onto fewer servers, while Instinct GPUs handle inference.

One potential use pairs locally hosted embedding models with SQL Server 2025’s native vector capabilities. The models would generate vectors that help applications retrieve information by meaning and use the results to ground AI responses. Shared inference services could also support coding agents and internal knowledge assistants.

AMD previously introduced a Helion enterprise rack design that supports up to 16 Instinct MI350P GPUs. The Windows Server support adds an operating system path for enterprise deployments.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…