Вештачка интелигенција

DriveNets Introduces Layer 2 Switching Solution for AI Training and Inference

Summary

DriveNets, a software development company specializing in distributed routers powered by Broadcom ASICs, has recently unveiled a new solution for AI training and inference. By incorporating Layer 2 switching into their network operating system (DNOS), DriveNets aims to deliver the […]

DriveNets Introduces Layer 2 Switching Solution for AI Training and Inference

DriveNets, a software development company specializing in distributed routers powered by Broadcom ASICs, has recently unveiled a new solution for AI training and inference. By incorporating Layer 2 switching into their network operating system (DNOS), DriveNets aims to deliver the necessary performance for handling AI workloads on Ethernet.

This move comes as a surprise, considering DriveNets’ previous statement that they would not venture into switching solutions. However, the potential that AI training and inference offers for demanding generative AI tasks has pushed DriveNets and other IT companies to find solutions that enable Ethernet to compete with Nvidia InfiniBand in terms of low latency, traffic shaping, and scalability.

In collaboration with Broadcom, DriveNets has developed Layer 2 switching in DNOS, utilizing Jericho 3-AI switch ASICs and Ramon 3 fabric elements. This distributed modular network architecture ensures efficient data transmission between network elements, with minimal latency and high bandwidth.

One of the key features of DriveNets DNOS is its use of “cell-based” communication, similar to the coherent memory protocol first used in NUMA servers in the late 1990s. This architecture enables deterministic performance and adaptive routing, which is crucial for telecom and service providers in their routing operations.

DriveNets plans to collaborate with network equipment manufacturers and implement Jericho 3-AI/Ramon 3 devices into a virtual switch that can support up to 32,000 GPUs, providing faster AI training execution compared to other Ethernet solutions. This scalability is inherent in the Jericho 3-AI/Ramon 3 design and represents a significant advancement in network technology.

In summary, DriveNets DNOS with Layer 2 switching allows Ethernet to compete with other technologies like InfiniBand in terms of performance, latency, and scalability, opening up new possibilities for AI training and inference.

FAQs