Microsoft Previews RDMA Over USB4 for Windows PCs
The developer preview lets multiple Windows PCs connected by USB4 cables split one AI model across their combined memory and GPUs.
Updated on Nov. 21, 2026 in Computers

Microsoft has announced a developer preview of RDMA over USB4 on Windows, which lets several PCs linked by USB4 cables run a single AI model together. The feature requires no RDMA network adapter, switch, or third-party driver.
Why it matters
Many of today's most capable open-weight AI models do not fit in one PC's memory, so pooling machines offers a way to run them locally. Low latency is essential because the machines must exchange data many times for every token generated.
Microsoft's preview measured 28 to 50 microseconds of round-trip latency for small messages between two PCs, about four times lower than Ethernet (UDP) over the same USB4 link, and three AMD Ryzen AI Max PCs with 128 GB of memory each ran Qwen3-235B-A22B at about 18.5 tokens per second at 4-bit quantization. Results come from Microsoft's own tests on clusters of up to three AMD and four Intel PCs, and performance on other configurations has not been independently reported.
The players
Microsoft
Microsoft is a multinational technology company that develops the Windows operating system and a wide range of software and cloud services.
AMD
AMD is a semiconductor company that designs processors and GPUs, and its Ryzen AI Max chips were used in Microsoft's testing.
Intel
Intel is a semiconductor company that designs processors, and its Core Ultra Series 3 PCs were used in Microsoft's testing.
The details
The feature lets Windows applications reach the USB4 link through NetworkDirect, so frameworks that already support that API need no USB4-specific code. Each PC needs N-1 USB4 ports to connect N machines, with a direct cable between each pair, and Microsoft has released the new NetworkDirect provider and a llama.cpp-based testbed as open source.
Timeline
The required Windows build is coming soon to Windows Insiders in the Experimental channel.
The Tech Race
RDMA has long been used to link servers in AI data centers over specialized high-speed networks, and this preview extends that approach to desktop-class Windows machines connected by USB4. It places Microsoft's Windows platform in the effort to run large AI models on local hardware rather than only in the cloud.
Users who own multiple USB4-equipped Windows PCs could eventually pool memory and GPUs to run larger AI models locally, once the feature reaches a stable release. Because it is off by default and requires a registry key and test signing, which means disabling Secure Boot, most users will not be able to use it without technical setup.
The takeaway
Pooling existing PCs over USB4 may let people run AI models too large for a single machine without buying new hardware. Anyone experimenting with the preview should weigh the security tradeoff of disabling Secure Boot.
Further reading
Explore more Computers coverage for context on how local hardware is handling larger AI workloads.
Source note: This article includes information reported by Teknoloji Haberleri - ShiftDelete.Net.



