Market Perspectives Guest Blog
Market Perspectives Guest Blog RSS FeedRSS

See what market experts predict for various markets related to SI, PI and EMI applications.

Signal Integrity

A Quantum Leap in Data Center Networking

With Help from Spirent, H3C Conducts the Industry’s Highest-Scale 800G Ethernet Test to Date

April 12, 2024

Artificial Intelligence-Generated Content (AIGC) is rewriting the rules for data center networks, driving a need for extreme throughput and density at unprecedented scales. To meet these requirements, hyperscalers need high-performance 800G network solutions as quickly as vendors can deliver them. 

Against this backdrop, H3C, a global leader in data center networking, recently achieved a major milestone by completing the industry’s first large-scale, high-density 800G Ethernet test for a next-generation switching platform. Working with test and assurance partner Spirent, H3C demonstrated the ability to deliver 64 ports of 800G traffic concurrently—51.2 Tbps total throughput—in a single platform, an industry first.  

The test validated the performance and reliability of H3C’s new flagship 800G platform, the S9827 co-packaged optics (CPO) silicon photonic switch. Just as important, it showed that when vendors and testing providers collaborate, they can accelerate the delivery of breakthrough networking solutions for AI and beyond. 

The Growing AI Challenge

In the world’s largest data centers, the rise of AIGC has unleashed a flood of large-scale data- and compute-intensive workloads, and an urgent need for higher-performing networks. AI model training in particular requires clusters that support high-speed data exchange among tens of thousands of processors executing billions of parallel computations. As AI models grow larger (with model parameters growing by 1,000x every three years, according to Dell’Oro Group), AI clusters will soon need to support workloads with trillions of parameters. 

These requirements are accelerating adoption of 800G (and soon, 1.6T) as the standard for large-scale AI training clusters. They also create an immediate demand for denser, higher-performing network platforms. As processor performance and computing efficiency continue to improve in the coming years, data center networks must improve as well or risk becoming bottlenecks that impede the growth of AI. 

Consider an operator building an AI cluster with 5,000 modern accelerators such as the NVIDIA H800 GPU, where each GPU corresponds to a 400G networking interface. Batch copying and transferring data among so many GPUs would quickly max out network resources in a conventional spine-leaf architecture, so the operator would use a 1:1 fat-tree architecture with no convergence. However, the scale of this architecture—the maximum total GPUs that can be connected—depends directly on the performance of each individual switch. 

Until now, the densest, highest-performing switches maxed out at 25.6 Tbps total throughput—effectively placing a hard limit of 2,000 GPUs for a two-layer architecture. Data center operators therefore have two options. They can adopt three-layer architectures (with many more devices, exponentially more complex topologies, and far higher costs), or they can increase throughput and density of each device. In this context, it’s easy to understand why hyperscalers prefer Option 2—and why networking vendors are racing to deliver it. 

Building a Next-Generation Switch

H3C developed its 800G switching platform to address these growing AIGC networking requirements. The S9827 is designed to deliver 64 ports of 800G at line rate, or 51.2 Tbps total throughput. This level of performance is unprecedented in a single platform, providing the ability to deploy AI training clusters with up to 8,000 accelerators in a two-layer architecture. 

At the same time, it’s not just H3C customers that require higher-scale next-generation technologies. To validate the performance of the S9827, H3C needed a new generation of testing capabilities as well. Testing solutions designed for 400G technology cannot support new QSFP-DD and OSFP interfaces, nor can they emulate the extreme data volumes, workload complexity, or traffic patterns of 800G AI clusters.  

For example, among many innovations introduced in next-generation platforms, 800G ports will be the first to use 112G serializer/deserializer technology, doubling data rates per channel. At these transmission speeds, even the slightest jitter in the testing environment can lead to errors and deviations. To validate the S9827 then, H3C needed a very high level of compatibility and stability across all 800G ports, optical modules, and testers. And they needed an entirely new testing platform capable of emulating high-scale AI workload traffic end to end. 

H3C used the new Spirent B2 800G appliance—the industry’s first high-density 800G OSFP and QSFP-DD test platform. The B2 test solution provides the density, flexibility, and performance to validate next-generation data center technologies against benchmark performance and scale standards, regardless of interconnect. Given the novelty of 800G technologies and the first-of-its-kind nature of the testing required, Spirent worked closely with H3C throughout the testing process.

Inside the Test

H3C used Spirent’s new 800G testing suite, including 800G full line-rate streaming tests, large and small packet forwarding capability tests, and latency tests, all under diverse network conditions. For these tests, they measured transmission performance in full mesh networking mode to provide a realistic simulation of all-reduce operations performed in a typical AI cluster. H3C also focused on port forwarding delay—the most important metric determining the transmission efficiency of AI training data. Here, the B2 testing solution validated a delay of approximately 1 µs for the S9827—in line with the highest-performing switches.

Ultimately, H3C validated that the S9827 delivers total switching capacity up to 51.2 T, with all 64 ports achieving 100% line-speed forwarding under different traffic patterns. Transmission rates for each port reached 800 Gbps—the first platform to ever achieve validated switching performance at this scale. The platform’s integrated CPO silicon photonics fully achieved the high throughput required for AIGC clusters and other emerging high-performance data center applications.  

The Spirent B2 800G appliance played an important role in this industry-first test, and in helping H3C deliver the reliable next-generation performance that their customers need and expect. For the first time, H3C engineers were able to observe all 800G ports on the S9827 with full traffic and validate that it could consistently complete every benchmark test.  

“Spirent helped us verify that the S9827 delivered the extremely high performance, density, and reliability we envisioned,” says Yutao Li, vice president of Network Product Line of H3C and general manager of Switch Product Line. “We were able to verify detailed metrics such as physical layer compatibility, transmission performance, transmission quality, failover, and more. This is essential, as identifying even minute flaws will help us improve our product and ensure that we can lead the way in integrating all the industry’s leading component suppliers into our solutions.” 

Looking Ahead

As AIGC and other next-generation applications continue to evolve, we can expect even greater demands for capacity, performance, and scale in data center infrastructures. Network suppliers like H3C will continue to push the boundaries of what’s possible as the industry adopts higher-speed technologies, from 800G Ethernet, to emerging 1.6 T standards and beyond. Testing partners like Spirent will also continue to provide the validation and emulation tools necessary to deliver next-generation products—even while next-generation standards are still being developed. Together, we can ensure that the entire data center ecosystem—from devices to semiconductors to optics to interconnects—will continue to meet the performance demands of tomorrow’s AI workloads. 

You must login or register in order to post a comment.