We’re excited to announce Guest RDMA (Preview) in Azure, beginning today in our UK South region, bringing high-throughput, ultra-low latency networking directly into guest virtual machines anywhere within the same region. By using Azure Boost to enable RDMA capabilities on our frontend network within the guest OS, applications can bypass the traditional networking stack and offload the transport protocol to the NIC - reducing CPU overhead and delivering consistent, high-performance communication for a variety of workloads including AI inference and training, storage, database, and HPC workloads.

RDMA 

RDMA is a networking technology that enables direct memory access between machines without involving the CPU or operating system kernel. Instead of passing data through the traditional TCP/IP stack, RDMA offloads data movement to the network interface card (NIC), allowing applications to read or write remote memory directly.  Azure uses RDMA extensively today at the infrastructure layer for storage, and in backend networks for AI and HPC scenarios. Extending RDMA support to guest VMs on the frontend network allows VMs to take advantage of these offloads for a broad range of scenarios within a region, including across Availability Zones. 

Key benefits of enabling RDMA for applications in your VMs include: 

• Kernel bypass and low latency: Data transfers avoid the OS networking stack, eliminating context switches and copies, reducing latency and jitter.

• High throughput: a hardware-based transport delivers up to 100Gb/s single-connection throughput depending on VM SKU.

• Low CPU usage: CPU overhead of the TCP/IP stack is eliminated since data movement is offloaded to hardware. 

Guest RDMA Scenarios

Applications that typically use Guest RDMA involve frequent or large data transfers, such as: 

• AI/ML training and inference – Guest RDMA supports GPU Direct RDMA, enabling direct data movement between GPUs across VMs with minimal CPU involvement. This is useful for doing training across GPUs that don't share a single backend network, disaggregating GPUs for inference, loading stored KV Caches, and other parts of AI workloads, improving throughput and latency. 
• Distributed storage systems and databases – Azure Boost supports both kernel mode RDMA (e.g., NFS, SMB) and user mode RDMA (e.g., shared memory and database workloads), delivering high throughput and low CPU utilization for storage and data platforms. A key property of Guest RDMA in Azure is that it allows systems distributed across multiple Availability Zones (AZs) within a region to communicate over high speed RDMA. 
• High Performance Computing (HPC) – RDMA delivers low latency, high bandwidth VM to VM communication, which is critical for tightly coupled, communication intensive HPC applications and MPI workloads. 

Enabling RDMA in Guest VMs

To enable Guest RDMA, create VMs using the following guidelines below: 

• Guest RDMA supports RDMA connections between VMs in a VNET talking directly to each other within the same region. The preview is currently available in the UK South region, with more regions to follow.
• Recommended Linux Supported Distribution: Ubuntu 24.04 LTS
• Required user-space packages:  >= rdma-core (50.0-2ubuntu0.2)
• Supported Kernel to use:   >=6.8.0 - 1044-azure, recommend 6.17 if available 

Supported VM sizes:

You can take advantage of Guest RDMA by utilizing the following virtual machine type

D-series: Dlsv6, Dldsv6, Dsv6, Ddsv6 

E-series and L-series: Esv6, Edsv6, Lsv4 

Network Optimized: Dnlsv6, Dnldsv6, Dsv6, Dndsv6, Esv6, Edsv6 (Public Preview) 

Remote Storage Optimized: Ebsv6, Ebdsv6 (Public Preview) 

FX: FXmsv2, FXmdsv2 

Limitation