Microsoft previews USB4-based RDMA for multi-PC local AI inference
11/20/2026 — 11/20, 17:20·1 sources (all)·1 reports (all)
Key takeaways
- Microsoft released a developer preview of a USB4-based RDMA feature that lets two or more Windows 11 PCs pool memory and GPU compute over a…
Background & analysis
On November 19, 2026, Microsoft published a blog post announcing a developer preview of an RDMA experience built on USB4, designed to let multiple Windows 11 PCs work together on a single AI model. According to an IT Home report published on November 20, Microsoft says two or more qualifying computers can be linked with USB4 cables without buying a dedicated RDMA network interface card or setting up a switch. Paired with a compatible inference framework, the machines can pool their memory and GPU resources to run one AI model together. The report explains that RDMA lets one computer read and write data in an authorized memory region of another efficiently, cutting the involvement of the operating system and CPU in data transfer, which yields low latency, high throughput, and modest CPU usage.
In Microsoft's published test figures, round-trip latency between the computers measured roughly 28 to 50 microseconds, about one quarter of what a traditional Ethernet UDP approach delivers over the same USB4 link when the amount of data sent in one transfer is small. In a preliminary test, Microsoft used three AMD mini PCs, each with 128GB unified memory, connected over 40Gbps USB4, running a 4-bit quantized Qwen3 model with 2350B parameters. Generation speed reached about 18.5 tokens per second, while AMD's own internal test recorded roughly 22 tokens per second. The solution remains a developer preview, and Microsoft has not yet announced a timeline for general availability.
AI-generated from 1 reports · updated 2 days ago
- Heat index
- 30
- All sources
- 1
- All reports
- 1
- First seen
- 2 days ago
Reports on this story headlines open the original
Microsoft released a developer preview of USB4-based RDMA that lets two or more Windows 11 PCs pool their memory and GPU power to run a single AI model, without dedicated RDMA NICs or a switch. In early tests, three AMD mini PCs with 128GB of unified memory each, linked over 40Gbps USB4, ran a 4-bit quantized 235B-parameter Qwen3 model at about 18.5 tokens per second.
IT之家AI score 57
Other stories people are talking about
- Heat index 1143RisingAnthropic Reaches $1.6 Trillion Market Value in First Trading Week After IPO21 sources
- Heat index 1011Anthropic Mathematician Says Claude Formalized the Generalized Riemann Hypothesis21 sources
- Heat index 622Anthropic IPO total raise reaches $93.5 billion after greenshoe exercise20 sources
- Heat index 596RisingThinking Machines Lab Open-Sources Inkling 2 Mini with 32GB Model Card Recommendation10 sources
- Heat index 490RisingSEC probes Nvidia accounting practices as shares post seven-week slide9 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

