Recently, as the traffic of AI service expansion users increases, the traffic and computational volume of AI data centers are increasing, and efficiency degradation and learning failure due to GPU delay at the network and interconnect stages have emerged as topics of discussion in the industry.

Keysight Technologies (hereafter referred to as Keysight) held a press conference at the Yangjae El Tower in Seoul on the 2nd and announced the official launch of the Keysight AI (KAI) Data Center Builder in Korea.

We have launched a new end-to-end solution that enables customers to expand AI processing capabilities within their data centers by emulating and validating AI cluster components with real workloads.

The KAI Data Center Builder comes with AresONE hardware, a high-density, high-performance Ethernet test solution with speeds up to 800G, and software solutions that can emulate AI model training workloads and integrate them into design and verification.

Deputy Director Kim Do-ho said that the network sector has emerged as a major issue in AI data centers, and that “in inference and learning, the network latency for communication between servers, server racks, and data centers is too long compared to the GPU’s computational latency, which has a significant impact on lowering GPU efficiency.”quo; he pointed out.

The industry has become a hot topic of discussion, with 62% of GPU latency being caused by communication factors rather than computation. In addition, 21% of the 43% of learning failures are caused by network issues, requiring a solution to this issue.

Accordingly, Keysight is introducing the KAI Data Center Builder, focusing on solving AI industry challenges with solutions that cover the entire area from the physical layer to the protocol layer and application layer, especially supporting the network and interconnect sectors.

KAI architecture supports AI suppliers and network equipment manufacturers to debug cutting-edge high-speed digital designs at the design stage and meet the latest △PCIe △DDR △CXL standards. It provides solutions to verify component specifications such as high-speed interconnects, cables, and chipsets at the development stage and to verify workload performance at the system level.

The distribution and operations team explained that by verifying and tuning the performance of the entire data center system, they can identify performance issues in advance before actual operation and reduce the risk of failure.

In addition, today's data centers are seeing a trend of developing switches with built-in optical signal conversion devices, centered around big tech, to optimize AI server performance and resolve heat generation issues. Developers of optical transceivers and related switches are conducting research to incorporate optical modulation into chips to solve the problems of securing space between server racks and heat circulation.

For these, Keysight provides the DCA-M sampling oscilloscope and INPT-1600GE interconnect and network performance tester, strengthening support for rapid construction of AI data centers, achievement of signal integrity requirements, and interconnect verification.

The DCA-M sampling oscilloscope supports up to 240 Gbps lane optical signal analysis and provides high optical measurement sensitivity with integrated clock recovery up to 120 Gbaud. It is designed to meet the optical test needs of 1.6T transceivers.

“The challenge today is that problems can arise in sending results from AI data centers to devices, and we don’t know where they might arise,” said Lee Sun-woo, CEO of Keysight Korea. “There are too many connections in the value chain from devices to the network, and too many points to inspect,” he said, emphasizing that these problems are particularly evident in scale-out and scale-up.

“KAI Data Center Builder provides data center solutions in four areas, covering overall network, interconnect, compute, and power, and we plan to strengthen this in the future,” added the representative.