In August, Keysight announced the release of the first public version of KAI Data Center Builder – a robust AI workload emulation solution that accelerates design and deployment of AI data center infrastructure through performance evaluation and benchmarking.

Introduced earlier this year as part of the Keysight AI (KAI) umbrella of solutions, this new product has already been used by many network equipment and silicon vendors, as well as hyperscale AI data center operators via a very successful beta preview program. KAI DC Builder is now broadly available for everyone to evaluate and order.

The 1.0 release centers around the Collective Benchmarks application, designed to characterize the performance of backend networks in AI clusters. These networks interconnect hundreds and thousands of GPU servers during AI model training and inference. AI workloads rely on the high bandwidth and low latency of backend networks to partition billions — or even trillions — of parameters of modern AI models across multiple GPUs, accelerate model training by running data through multiple instances of the same model, and exchange data between multiple expert models.

All these techniques rely on data movement patterns called collective operations, which vary in type, scale, algorithms and amount of data. In the Collective Benchmarks application, Keysight has adopted and enhanced industry best practices to measure performance of such operations over the backend network.

KAI Data Center Builder can run the Collective Benchmarks using either Keysight AresONE hardware traffic generators, or x86 servers with Remote Direct Memory Access Network Interface Cards (RDMA NICs). This allows customers to validate and compare performance of network infrastructure for AI clusters directly — without depending on the broader hardware and software stack within the GPU servers. While use of GPU servers for benchmarking is important to understand end-to-end performance, it leaves low-level aspects of individual data and packet movements hidden behind layers of software and hardware. This makes it difficult to understand root causes of low performance. In contrast, running the benchmarking directly from the AresONE hardware test ports connected to the network fabric exposes any deficiency and limitation the network has. KAI Data Center Builder enables users to drill down into such low-level measurements, providing deeper performance analysis.

Once network performance becomes satisfactory, moving benchmarking from the AresONE hardware to servers allows customers to focus on understanding the impact RDMA NICs have on overall performance. With KAI DC Builder, users can compare measurements taken from AresONE hardware and RDMA NICs to identify the difference and design a plan for mitigation.

Finally, the Collective Benchmarks application also facilitates running end-to-end benchmarks with GPUs in the loop by orchestrating execution of NVIDIA’s nccl-tests* utility. Comparing results taken with nccl-tests and RDMA NICs or AresONE closes the validation cycle to confirm improvements.

NOTE*: nccl-tests are developed and maintained by NVIDIA Corporation.

Beyond v1.0

Version 1.0 is designated as an Early Adopter (EA) release, reflecting Keysight’s rapid pace in developing new capabilities within the solution. You can try KAI Data Center Builder today and expect more from the Keysight team!