Accelerating AI Training and Inference with RoCE Networking

With lower deployment costs and high-performance lossless networking, RoCE is emerging as a scalable alternative to InfiniBand for AI GPU clusters. Dilight delivers validated RoCE interconnect products and end-to-end solutions, fully compatible with the NVIDIA Ethernet ecosystem and seamlessly integrated with Spectrum switches and ConnectX adapters, accelerating the scaling and deployment of AI training and inference infrastructure.

Why RoCE for AI Networking

RoCE for Diverse AI Cluster Requirements

RoCE supports AI cluster deployments from mid-size to ultra-scale, delivering stable high-performance through scalable Ethernet architecture, flexible topology design, and efficient congestion control, while its open ecosystem and routable capabilities ensure smooth network evolution to meet future compute growth and diverse workloads.

RoCE for AI Training, Inference, and HPC

RoCE networks are widely used in AI training and inference, ML, HPC, and distributed storage. By leveraging RDMA, lossless Ethernet, and hardware offloading, it enables low-latency, high-throughput communication between GPUs, compute nodes, and storage systems. RoCE delivers near-InfiniBand performance while retaining the scalability and cost efficiency of Ethernet.

Key Benefits of RoCE-Based Ethernet Fabrics

Reuse Existing Infrastructure and Lower Costs

High-performance networks traditionally rely on InfiniBand, which involves expensive switches and NICs, a closed ecosystem, and dedicated operations teams. RoCE leverages existing Ethernet infrastructure (with PFC/ECN support) without large-scale network upgrades, significantly reducing hardware costs. Its open ecosystem enables interoperability across NVIDIA, Arista, Broadcom, Cisco, and other vendor platforms, reducing procurement and upgrade costs.

Optical Interconnect in an Open RoCE Ecosystem

NVIDIA builds a fully optimized end-to-end RoCE ecosystem for AI workloads, spanning GPU compute, NICs, switches, and software layers.

While NVIDIA defines end-to-end system optimization, the actual performance of RoCE networks depends on the physical interconnect layer that connects each component. Optical interconnect becomes the critical foundation for sustaining system-level performance.

On top of the optical interconnect layer within the NVIDIA RoCE ecosystem, Dilight delivers end-to-end interconnect solutions by integrating key components—such as optical transceivers and AOC/DAC cables—into mainstream AI network architectures, enabling highly reliable, ultra low-latency, high-bandwidth connectivity between switches and servers.

Ensures end-to-end bandwidth efficiency

Maintains GPU cluster latency consistency

Guarantees lossless transmission reliability

Dilight RoCE Solutions for NVIDIA Ethernet Platforms

Dilight high-speed optical interconnect solutions are deeply optimized for NVIDIA Ethernet platforms, targeting RoCE-based AI and HPC deployments, and providing critical physical connectivity for large-scale GPU clusters.

Deeply compatible with NVIDIA ConnectX-series NICs/Ethernet SuperNICs and Spectrum Ethernet switches, enabling end-to-end RoCE network architecture.

Provides a full range of 400G/800G/1.6T optical transceivers, DACs, and AOCs, serving as the high-speed physical link layer between servers and switches.

Designed and validated for RoCEv2 and AI training traffic patterns, ensuring reliable operation in Spectrum-based lossless Ethernet environments, and delivering low latency, high throughput, and long-term stability.

With strong compatibility across NVIDIA Ethernet platforms, Dilight helps build scalable, high-performance optical interconnect infrastructure for next-generation AI networking architectures.

800G/400G RoCE Solutions for NVIDIA DGX Blackwell (B300/B200)

Spine-Leaf

Leaf-Server

Dilight Products for Scalable AI Cluster Interconnect

AI Data Center Switches

Based on Broadcom Tomahawk 5/4 & NVIDIA Spectrum-4 chip, Dilight 800G/400G/200G AI data center switches are preloaded with ONIE, SONiC, NDD OS, which offers reliable and high-performance networking solutions for AIDC and cloud.

Learn More

NVIDIA Ethernet Switches

NVIDIA Spectrum-3/4 Ethernet switches enable high-performance, low-latency, lossless Ethernet fabrics for AI and cloud-scale data centers.

Learn More

Networking Adapters

Server-side high-speed adapters such as ConnectX-7/8/9, supporting RDMA/RoCE to accelerate AI training, cloud, and data center.

Learn More

NVIDIA DPUs and SuperNICs

NVIDIA DPUs (BlueField-3) and SuperNICs, enabling infrastructure offload, secure high-performance networking, and ultra low-latency connectivity for AI and HPC workloads.

Learn More

Ethernet Transceivers

Full range of optical transceivers from 100G to 1.6T, available in OSFP, QSFP-DD, QSFP112, QSFP56, and QSFP28 form factors, enabling flexible deployment across diverse AI network architectures and applications.

Learn More

Fiber Patch Cables

Optimize AIDC network cabling architecture, enhance network stability and operational efficiency, and enable high-density, scalable, and reliable network deployments.

Learn More

AOC Cables

Integrated optical links for short-reach in-rack interconnects, plug-and-play, low power consumption, and simplified deployment for high-density GPU clusters.

Learn More

DAC Cables

Passive, ultra low-latency, low-power copper interconnects for in-rack and adjacent-rack connections, delivering cost-efficient (low) TCO.

Learn More

Why Dilight RoCE Products?

High-Performance Switches and Optics for AIDC

Dilight's AI switches and optics are designed for peak performance in demanding AI environments. Our 800G/400G 51.2T switches, powered by the Broadcom Tomahawk 5/4 or NVIDIA Spectrum-4 chip and pre-installed with the SONiC network OS, support RoCEv2 lossless networking, with full Layer 2 and Layer 3 forwarding capabilities. Cognitive routing, such as global load balancing, reactive path rebalancing, hardware link failover, and drop congestion notification, optimizes network efficiency and imporve AI/ML training performance.

For optics, Dilight integrates advanced thermal management solutions to minimize network outages in AI training clusters. Utilizing cutting-edge DSP (digital signal processing) and algorithm tuning, our RoCE transceivers achieve a pre-FEC bit error rate (BER) between 1E-8 and 1E-10, with near-zero post-FEC errors. These advancements guarantee exceptional performance, robust reliability, and seamless connectivity for data centers, AI/ML clusters, and HPC environments.

Proven RoCE Compatibility in Real AIDC Test Environments

Dilight Test Center builds real deployment environments based on equipment from leading vendors such as NVIDIA, Arista, Broadcom, and Cisco, providing deep compatibility coverage across 140+ vendor ecosystems. Through rigorous validation on core platforms—including NVIDIA SN5000, Arista 7060X6 and Cisco Nexus series switches, as well as NVIDIA ConnectX-8 and Broadcom BCM957608 Thor 2 NICs—Dilight ensures that RoCE product passes 100% interoperability and performance compatibility testing. Detailed test reports are available upon request. This thorough compatibility testing ensures smooth interoperability and high performance, whether deployed in single-brand or multi-brand environments.

Request Test Reports arrow-right

100% Tested in NVIDIA / Arista / Cisco and Others

Arista 7060X6/5 Series

7060X6-64PE7060X6-32PE7060PX5-64E7060DX5-64S7060DX5-32

Cisco Nexus 3400/9300 Series

N3K-C3432D-SN9364E-SG2-QN9364E-SG2-O9364D-GX2A9348D-GX2A

RoCE Networking Design & Deployment Capabilities

Validated RoCE Networking Design Expertise

Dilight focuses on mature RoCE interconnect design and configuration validation, embedding R&D expertise and large-scale deployment experience directly into product portfolio. Our switches, transceivers, AOC/DAC cables, and NICs are optimized for AI and HPC workloads, with validated interoperability across multi-vendor RoCE environments.

Beyond hardware, Dilight delivers deployment-proven port configurations, speed combinations, and network deployment solutions, significantly reducing network design complexity and minimizing tuning and scaling risks.

End-to-End RoCE Deployment Support

Optics-to-Device Compatibility Matrix

Quickly identify compatible transceivers and cabling solutions for your network devices. Built on extensive multi-vendor interoperability testing, the matrix helps simplify deployment and ensure stable, verified compatibility across AI Data Center, Cloud, and HPC networks.

View Compatibility Matrix

From Design to Deployment

Explore our AIDC Service for end-to-end support covering design, deployment, optimization, and delivery assurance for AI infrastructure.

Design & Technical Support

Deployment & Remote Optimization

Delivery & Quality Assurance

Contact Us

Ready to Scale Your AI Cluster with Dilight RoCE Product? Talk to Our RoCE Networking Specialists Today.

First Name*

Last Name*

Email Address*

Company/Organization

Country/Region*

Phone Number*

+852

How Can Dilight Optimize Your RoCE Interconnect Architecture?*

I agree to Dilight's Privacy Policy and Term of Use.

Submit

Contact Us

Ready to Scale Your AI Cluster with Dilight RoCE Product? Talk to Our RoCE Networking Specialists Today.

First Name*

Last Name*

Email Address*

Company/Organization

Country/Region*

Phone Number*

+852

How Can Dilight Optimize Your RoCE Interconnect Architecture?*

I agree to Dilight's Privacy Policy and Term of Use.

Submit

Loading content… / 正在加载内容…

Accelerating AI Training and Inference with RoCE Networking