In high-performance embedded computing (HPC) applications, the ability to efficiently network data to a GPU over Ethernet is becoming increasingly crucial. As sensor-based platforms become more reliant on Ethernet communications for various payloads, the demand for robust, low-latency data transfer solutions grows. RDMA over Converged Ethernet (RoCE) offers numerous advantages in these settings, providing low latency, high throughput, and reduced CPU overhead, all of which are essential for real-time operations.
Remote Direct Memory Access (RDMA) is a technology that allows devices to directly access the memory of remote devices without involving the CPU. When RDMA is used in conjunction with GPUs, it enables efficient data transfer between GPUs and other devices in the HPEC, such as network adapters or storage systems.
RDMA over Converged Ethernet is a networking protocol for using RDMA over an Ethernet connection. RoCE enables direct memory access from the network interface to the GPU, bypassing the CPU and thereby reducing latency and freeing up CPU resources for critical tasks. This is particularly beneficial for applications such as video encoding/decoding, computer vision, autonomous navigation, and artificial intelligence (AI), where timely and efficient data processing is paramount. Additionally, the inherent scalability and compatibility of RoCE with existing Ethernet infrastructures make it a cost-effective and flexible solution, facilitating seamless integration and future-proofing investments.
Unlike traditional transport protocols such as TCP/IP, where packets or messages are sent across a network, RoCE enables network interfaces to directly write to or read from memory locations. This bypasses the CPU, eliminating the need to copy packets into and out of a kernel-level buffer.
Because adopting this approach requires a fundamental shift in network architecture, different software is needed to take advantage of this technology.
RoCE has three main benefits: a reduction in latency, a reduction in CPU usage, and an increase in network bandwidth. RoCE enables high data transfer rates and low latency communication, which are essential for processing the large volumes of data generated in HPEC applications. As network sizes grow, RoCE maintains performance by reducing the load on CPUs and enabling direct memory access.
In a recent test performed by EIZO’s in-house engineers, standard TCP/IP networks achieved roughly 30-40 gigabits per second. Switching to RoCE on the same hardware yielded speeds of 100 gigabits per second. This increase is highly impactful to the C5ISR space, where massive amounts of data need to be transmitted rapidly at the edge.
Much of this bandwidth improvement is due to the CPU no longer needing to handle packet copying or TCP processing. During the experiment, the TCP/IP bandwidth test consumed three cores on the receiving device, while the RoCE bandwidth test used minimal CPU resources despite utilizing more than twice the bandwidth.
Additionally, with the CPU freed from constant buffer copying, packet transmission speeds improved significantly. We observed approximately a 90% reduction in transmission latency with RoCE compared to TCP/IP. For an in-depth analysis of the technical evaluation performed, read our latest white paper titled “Unlocking Lower Latency and Higher Bandwidth with RDMA/RoCE on NVIDIA Jetson AGX Orin.”.
By leveraging the advanced capabilities of RoCE, embedded edge systems can achieve significant improvements in data processing speeds and overall system performance, driving more effective and timely decision-making in the field.
Edge platforms require robust and flexible network infrastructures to support real-time data processing and communications. Given the existing investments in Ethernet-based networks within dynamic environments, it is crucial to implement advanced technologies like RoCE without requiring significant hardware overhauls. RoCE requires specific drivers and software support, which can be installed on existing systems. This software-based approach simplifies the transition and enables seamless integration into the current network architecture.
As high-performance embedded (HPE) systems evolve, the volume and complexity of data they handle will continue to grow. RoCE’s scalability ensures that HPE systems can accommodate increasing data demands without compromising performance or reliability. By providing a robust, high-performance data transfer solution, RoCE enables systems to process and analyze large datasets efficiently, facilitating timely and informed decision-making in mission-critical scenarios.
In summary, RoCE’s ability to scale in larger network environments and support growing data demands makes it an ideal solution for modern applications. Its combination of high throughput, low latency, and efficient resource utilization ensures that edge operations can keep pace with technological advancements and evolving mission requirements.
EIZO Rugged Solutions has two offerings that support RoCE via NVIDIA ConnectX NICs.
The Condor™ AGX-ORIN64-HPC is an OpenVPX 3U form factor single board computer designed with the NVIDIA Jetson AGX Orin 64GB (Industrial) SoM that hosts an NVIDIA Ampere architecture GPU and a 12-core Arm® Cortex®-A78AE CPU. The Condor AGX-ORIN64-HPC incorporates the NVIDIA ConnectX®-7 NIC, enabling high-speed data transfer of up to 100Gb/s and 25Gb/s using Gigabit Ethernet. This, along with NVIDIA GPUDirect® RDMA over converged Ethernet (RoCE), enables high-speed communication between distributed sensors, RF transceivers, and platforms.
The Condor™ GR2S-A4500-ETH is a rugged OpenVPX 3U form factor video graphics and GPGPU card based on NVIDIA® Ampere™ A4500 GPU. The Condor GR2S-A4500-ETH is designed with an NVIDIA® ConnectX®-7 SmartNIC which enables high-speed data transfer and enhanced storage performance with NVIDIA GPUDirect™ RDMA and RoCE. The ConnectX-7 fabric includes an embedded PCIe Switch along with SmartNIC function, this allows for better board function and reduces the overall power footprint (lower SWaP) of the card. The card supports 40/100Gb Ethernet DataPlane interface and is compatible with OpenVPX VITA connectors. This solution offers systems with lower SWaP (Size, Weight, and Power) as the integration of the SmartNIC and PCIe Switch on a single module reduces the overall power footprint, making it more efficient and suitable for space-constrained environments.
Both the Condor AGX-ORIN64-HPC and Condor GR2S-A4500-ETH demonstrate EIZO Rugged Solutions’ commitment to providing state-of-the-art technology for high-performance embedded computing. By leveraging RoCE via NVIDIA ConnectX NICs, these products ensure high-throughput data transfer and efficient resource utilization, making them ideal for demanding C5ISR applications.
Download our capabilities overview to read how EIZO is uniquely positioned to help implement CJADC2 with broad capability for supporting mission-critical data.
EIZO Rugged Solutions is a pioneer in providing video, graphics, encoding, and recording solutions for over three decades.
