Newshot

Microsoft Unveils USB4-Based Multi-Windows 11 PC AI Collaboration, Stacking Memory and GPU Compute Like Building Blocks to Run Local Models

AI summary

Microsoft on Nov. 19 published a blog post announcing a developer preview of a USB4-based remote direct memory access (RDMA) experience that lets multiple Windows 11 PCs pool their resources to run AI models together. With two or more eligible PCs connected by USB4 cables, no dedicated RDMA network card or switch is required; compatible inference frameworks can draw on each machine's memory and GPU compute, with round-trip communication latency of about 28 to 50 microseconds for small transfers — a quarter of that of traditional Ethernet UDP on the same USB4 link. In initial tests, three AMD mini PCs with 128GB of unified memory each, linked over 40Gbps USB4, ran a 4-bit quantized 235-billion-parameter Qwen3 model at about 18.5 tokens per second (AMD's own testing: about 22), while four Intel Core Ultra Series 3 PCs with 64GB each ran GPT-OSS-120B at 54.2 tokens per second at roughly 1,000-token context.

Translation(model translation, not the source's own text)

Microsoft announces USB4 collaborative AI solution for multiple Win11 PCs, stacking memory / GPU compute like building blocks to run local models

IT之家, November 20: Microsoft yesterday (November 19) published a blog post announcing a USB4-based remote direct memory access (RDMA) experience in developer preview, enabling multiple Windows 11 PCs to jointly run AI models.

IT之家 note: RDMA allows one computer to efficiently read from and write to authorized memory regions on another computer, reducing OS and CPU involvement in data movement. Its main advantages are low latency, high throughput, and lower CPU usage.

According to the official description, after two or more compatible PCs are connected via USB4 cables, there is no need to buy dedicated RDMA NICs or configure switches; with compatible inference frameworks, they can pool their respective memory and GPU compute to jointly run a single AI model.

In terms of latency, Microsoft's tests show that for small single-transfer payloads, the round-trip latency of inter-PC communication is about 28 to 50 microseconds, only one quarter that of traditional Ethernet UDP over the same USB4 link.

In preliminary tests, Microsoft used three AMD mini PCs each with 128GB unified memory, connected via 40Gbps USB4, running a 4-bit quantized 235-billion-parameter Qwen3 model, achieving a generation

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗