OpenAI, Microsoft And Friends Build A Better, More Scalable Ethernet
AI-summarised brief · reviewed before publication
OpenAI, Microsoft, and other companies have developed a new network protocol called Multipath Reliable Connection, which enhances Ethernet scalability. The protocol increases the number of links between devices, reducing latency and cost. It achieves this by utilizing the same aggregate bandwidth of a switch ASIC to boost the number of links, rather than pursuing higher bandwidth ports. The MRC protocol is an extension of the current RDMA over Converged Ethernet protocol and incorporates features like adaptive load balancing and packet trimming. This approach enables the network to heal around link failures, ensuring uninterrupted AI training jobs. The development is significant as it addresses the need for scalable and efficient networks in AI systems.
💡 Why It Matters
- · By rethinking traditional network architecture, the MRC protocol offers a more efficient and cost-effective solution for scaling out AI systems.
- · It enables the creation of flatter networks with lower latency and power consumption, making it an attractive option for hyperscalers and cloud builders.
