Next-generation networks will need more bandwidth for Ethernet switching than ever before to support emerging, demanding workloads like artificial intelligence (AI) training and inference.

It's a challenge Keysight Technologies is looking to help solve with new testing capabilities to help validate the switching hardware required for next-generation networking. The new testing benchmark was done as a collaboration between Keysight and silicon vendor Marvell.

Using Keysight's AresONE-M 800GE test equipment and Marvell's Teralynx 10 Ethernet programmable Switch, the partners processed 51.2 Terabits per second (Tbps) of traffic across 64 links of 800 Gigabit Ethernet while measuring key performance metrics.

According to Keysight, the benchmark establishes a new standard for silicon validation of  Ethernet switching performance. It will enable network equipment manufacturers, chipset vendors and data center operators to ensure their solutions can handle the massive data flows required by advanced AI workloads.

While the initial benchmark was conducted with silicon from Marvell, the testing infrastructure approach is not limited to just that one vendor.

“This test is applicable for any high speed switch silicon that will be used in a data center fabric hosting AI applications,” Avik Bhattacharya, Senior Product Manager at Keysight, told SDxCentral. “The silicon needs to be able to forward traffic at 800GE full line rate across all 64-ports simultaneously and provide low latency, which is key to achieving the effective job completion time (JCT) for AI applications.”

The new benchmark that validates Switch silicon to 51.2 Tbps marks a significant performance milestone for networking Switch performance.

Bhattacharya said that the previous generation of data center Switch silicon supported 25.6 Tbps based on 64x400GE interfaces. The new generation of 800GE Switch silicon supports 64x800GE interfaces, totalling 51.2Tbps.

800GE electrical lanes are twice as fast as the 400GE electrical lanes. Also, he emphasized that it is two-times the total bandwidth and throughput, which is how much data is actually delivered over the network infrastructure, compared to the previous 400GE generation of switches.

The 800GE per interface speed or 51.2 Tbps technology is dependent on multiple components.

Among the new components is the 112 SerDes (serializer/deserializer) and new QSFP-DD800 or OSFP800 interconnect specifications.  Bhattacharya said that Keysight’s engineering team collaborated with world leading vendors and conducted early-stage interoperability testing with industry leaders in silicon, optical transceiver and network equipment manufacturing.

The AresONE-M test hardware built by Keysight is a new product, supporting the 8 lanes of data that is used to generate 800 Gbps of traffic per interface. Bhattacharya  said that the AresONE-M is a completely new technology compared to the previous generation SerDes, which was used to generate 400Gbps of traffic per interface.

Having 51.2 Tbps of switching available can be a big help to bandwidth intensive workloads like AI. But it's not just raw bandwidth alone that matters.

One of the critical measures of an AI compute cluster performance is JCT (job completion time), where a shorter  job completion time (JCT) helps in better utilization of the cluster.

“The switching silicon that connects the high performance GPUs can define the JCT for an AI application,” Bhattacharya said. “So it is extremely important to benchmark the low latency performance through the switches for achieving an effective JCT.”