RDMA and RoCE Overview

What is RDMA?

Remote Direct Memory Access (RDMA) is a technology that allows computers to quickly send data from one computer's memory to another without using the main processor.  This makes data transfer faster because the processor doesn't need to handle the work. RDMA is often used in places where speed is important, like data centres or cloud services.

One big advantage of RDMA is that it reduces the strain on the processor, letting it focus on other tasks. It also lowers the delay in sending and receiving data, and uses less power. This is especially helpful for applications that need to move a lot of data quickly, like real-time analytics or large databases.

What is RoCE?

RDMA over Converged Ethernet (RoCE) allows RDMA to work over regular Ethernet networks. Normally, RDMA needs special high-speed networks, but RoCE lets it use existing Ethernet setups, making it cheaper and easier to implement while still boosting data transfer speed.

There are two versions of RoCE: RoCE v1 and RoCE v2. RoCE v1 works within local networks, while RoCE v2 can work over larger networks, including the internet. Both versions help reduce delay and improve performance, making RoCE great for data centres, cloud services, and other high-speed communication needs.

NVIDIA GPUDirect RDMA

NVIDIA GPUDirect RDMA with RoCE is a technology that allows GPUs to directly exchange data over Ethernet networks without involving the CPU. This enables faster data transfers between GPUs in different systems, which is crucial for high-performance applications like AI, machine learning, and data analytics.

The new Condor™ AGX-ORIN64-HPC from EIZO Rugged leverages NVIDIA GPUDirect® RDMA over converged Ethernet (RoCE), enables high-speed communication between distributed sensors, RF transceivers, and platforms.
 

The Condor AGX-ORIN64-HPC uses the NVIDIA Jetson AGX Orin 64GB system-on-module (SoM) that hosts an NVIDIA Ampere architecture GPU and a 12-core Arm® Cortex®-A78AE CPU. The module’s 64GB of LPDDR5 system memory allows up to 205 GB/s memory bandwidth shared between the internal CPU, GPU, and accelerator engines. The NVIDIA Jetson AGX Orin features the NVIDIA Ampere architecture with 2048 CUDA® cores and 64 Tensor Cores for AI-accelerated performance and includes NVIDIA Deep Learning Accelerators for complex machine learning (ML) operations.

The Condor AGX-ORIN64-HPC’s modular design supports VITA 46/65 and the SOSA technical standards slot profile 14.2.16. The product is designed to meet MIL-STD-810 requirements for temperature, shock, vibration, and humidity.

For more information on RDMA and RoCE Overview talk to Sarsen Technology Ltd