Friday, November 20th 2026

Microsoft Brings RDMA over USB4 to Windows 11 for Daisy-Chained PCs

Microsoft is finally enabling Remote Direct Memory Access (RDMA) over the USB4 protocol on Windows 11, allowing multiple PCs to be daisy-chained and used together to process large workloads. This means that Windows 11 now has a mechanism to distribute a workload across multiple interconnected PCs, enabling faster processing. Essentially, AI inferencing at home is now limited by the number of PCs and their specifications, rather than a single PC. RDMA allows the workload to directly access system memory on another PC, all through USB4, which offers up to 80 Gbps bandwidth for fast data transfer.

Instead of using Ethernet over USB4 in a UDP packet configuration, Microsoft chose to implement RDMA over USB4 to significantly reduce data access latency. Microsoft interconnected two PCs to measure round-trip latency by message size, finding that RDMA over USB4 achieves latency between 28 and 50 microseconds. In contrast, Ethernet over USB4 using UDP results in latency between 116 and 194 microseconds, making the RDMA implementation roughly four times faster. This lower latency will greatly benefit systems running large workloads, especially in latency-sensitive tasks like AI inferencing. Below DGX Spark mini PCs are connected over Ethernet, not USB4, but it serves the purpose of illustration.
RDMA over USB4 builds on Windows 11's existing host-to-host USB4 networking to transfer data directly between the registered memory regions of two PCs. Each PC forms a separate USB4 domain, and the Windows connection manager establishes a peer-to-peer tunnel between them. A software RDMA provider maps send and receive queues and their completions to the USB4 host router's DMA rings, avoiding extra data copies. Each link uses one pair of rings for data and another for control traffic, so the available rings and ports on each PC limit the cluster size. Compatible frameworks access the link through NetworkDirect without requiring USB4-specific code. The feature requires Windows-managed USB4 ports with Kernel DMA Protection.
Source: Microsoft
Add your own comment

11 Comments on Microsoft Brings RDMA over USB4 to Windows 11 for Daisy-Chained PCs

#1
r00tcz
Finally! I hope it's baked into Windows USB stack and not something specific to ThunderBolt or Intel only. And will there finally be support for SMB Direct over RDMA for both Server and Client on Windows?
Posted on Reply
#2
domih
RDMA bypasses the multiple buffer data copies happening through the IP stack, so no wonder it beats IP on latency. This is a good news for users doing inference on Windows 11 on several $$$ AI PC. Probably, a market reserved to B2B customers and data labs. I believe most individual hobbyists do not have the kind of money for 2, 5, 10 or more Spark-based PC. Except the ones who won at the stock ownership lottery.
BuddyspikeThat’s actually great news for local LLM inference setups and along with technologies like DirectStorage for streaming weights to VRAM from RAID0 Nvme array it makes Windows 11 viable option again. AFAIK Linux doesn’t have good alternatives for these yet . I know they are being developed e.g. there’s USB4STREAM in Linux 7.3 , but it’s not used by any inference engine. I think Tinygrad is working on something similar to DirectStorage
On the Linux side, an individual hobbyist has been able to use, for years now, RDMA on used Mellanox VPI cards from eBay at 56gbps speed , using InfiniBand or IPoIB. Plus, now that most data centers moved or are moving to 200GbE, 400GbE, 800GbE, the Mellanox 100Gbps cards price has been going down, almost OK for individual hobbyists to buy. The real issue, for inference, is the local memory speed and latency. So you want to use a modern system with the fastest DDR5 you can find. Given the scarcity and prices of these DIMM, plan for a monster $$$$ budget. A no-no for an individual hobbyist these days.
Posted on Reply
#3
Buddyspike
That’s actually great news for local LLM inference setups and along with technologies like DirectStorage for streaming weights to VRAM from RAID0 Nvme array it makes Windows 11 viable option again. AFAIK Linux doesn’t have good alternatives for these yet . I know they are being developed e.g. there’s USB4STREAM in Linux 7.3 , but it’s not used by any inference engine. I think Tinygrad is working on something similar to DirectStorage
Posted on Reply
#4
dgianstefani
TPU Proofreader
Very interesting as this is much faster than the 10 Gbps ethernet most of these systems are capped at.
Posted on Reply
#5
Darmok N Jalad
Macs are able to do this over Thunderbolt, and it has made the high-end Mac Studios very desirable for AI workloads. From what I gather, 4 Studios interconnected gets you 3x the performance of one.
Posted on Reply
#6
dgianstefani
TPU Proofreader
Darmok N JaladMacs are able to do this over Thunderbolt, and it has made the high-end Mac Studios very desirable for AI workloads. From what I gather, 4 Studios interconnected gets you 3x the performance of one.
The Macs have Thunderbolt 5 too, which is more flexible bandwidth wise. Very nice.
Posted on Reply
#7
Buddyspike
dgianstefaniThe Macs have Thunderbolt 5 too, which is more flexible bandwidth wise. Very nice.
Yeah but even TB5 with RDMA it’s too slow for TP with larger models.
Posted on Reply
#8
TechLurker
I wonder if this will be of any use in slaving a few DDR2 and DDR3 PCs to a DDR4 or DDR5 for more RAM; or even more hilariously, the reverse; an old DDR2 PC with a USB-C add-in card borrowing memory off a DDR4 or DDR5 PC. Probably not, but it'd be hilarious if it were possible. Just borrow more RAM off idle PCs for slower/lower priority tasks while using the main system RAM for gaming/editing/CAD/AI/etc.
Posted on Reply
#9
mashie
Since you linked the DGX Sparks in the article, this is the write latency seen there while going across a switch:

192.168.10.2:rocep1s0f0 -> 192.168.10.3:rocep1s0f0 (10.0.0.0/24) - lat 2.67 us, bw 97.9 Gb/s, of 100 Gb/s
192.168.10.2:roceP2p1s0f1 -> 192.168.10.3:roceP2p1s0f1 (10.0.1.0/24) - lat 2.68 us, bw 97.9 Gb/s, of 100 Gb/s
192.168.10.3:rocep1s0f0 -> 192.168.10.2:rocep1s0f0 (10.0.0.0/24) - lat 2.66 us, bw 97.9 Gb/s, of 100 Gb/s
192.168.10.3:roceP2p1s0f1 -> 192.168.10.2:roceP2p1s0f1 (10.0.1.0/24) - lat 2.70 us, bw 97.9 Gb/s, of 100 Gb/s
Posted on Reply
#10
_roman_
I read it that way, USB 4 is / was still not fully / partly supported on Windows 11, although USB 4 is very likely sold for over 5 years. At least AM5 had usb4 common support. Intel likes their thunderbolt too much and hardly shipped with any usb 4 support, or it was not noticeable.

The core information missing is, since when USB 4 exists, since when that protocol exists and if it is an add on protocol later specified, etc.

USB 4 a hype with no software support regardless of operating system. What else is missing? not proper implemented? How does it relaetes to thunderbolt? Did thunderbold also not have that feature in windows ?
Posted on Reply
#11
dgianstefani
TPU Proofreader
_roman_I read it that way, USB 4 is / was still not fully / partly supported on Windows 11, although USB 4 is very likely sold for over 5 years. At least AM5 had usb4 common support. Intel likes their thunderbolt too much and hardly shipped with any usb 4 support, or it was not noticeable.
Thunderbolt support is USB4 with additional certification and a slightly higher standard, rather USB4 was created by slightly downgrading TB 4 eg pd, bandwidth distribution etc. So if it has TB 4 support it 100% has USB4 support.
_roman_USB 4 a hype with no software support regardless of operating system. What else is missing? not proper implemented? How does it relaetes to thunderbolt? Did thunderbold also not have that feature in windows ?
Again, TB has had this type of feature for quite a while, this is a more open and less strenuous standard that has come afterwards.

USB4 V2 is the same type of relationship with TB 5. 80 Gbps, TB 5 has option for 120/40 instead of 80/80 duplex.

Posted on Reply
Add your own comment
Nov 21st, 2026 08:32 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts