The Scale Up Ethernet SUE Framework for AI ML Accelerators скачать в хорошем качестве

The Scale Up Ethernet SUE Framework for AI ML Accelerators

Mohan Kalkunte - , Hugh Holbrook Chief Development Officer - Arista Networks AI-ML workloads in Scale-Up networks require low latency, high bandwidth, lossless communication across tightly-coupled XPU clusters. Traditional network interfaces are area-intensive and limit scalability. The OCP-contributed Scale-Up Ethernet (SUE) Framework is an Ethernet-based AI-ML architecture designed for efficient implementation. SUE standardizes a memory semantics XPU interface and enables a lightweight transport, leveraging AI Forwarding Headers (AFH), Credit-Based Flow Control (CBFC), and Link Layer Retry (LLR) to ensure deterministic performance and low tail latency. Aligned with OCPs mission of open infrastructure, SUE enables scale-up clusters to use merchant Ethernet silicon and open protocols. SUE is seamlessly supported by open OSes such as SONiC. In this session, we present how the SUE framework enables efficient scale-up interconnects for AI-ML clusters, simplifies deployment, and advances the vision of open, Ethernet-based Scale-Up AI.