Tesla Transport Protocol over Ethernet

lastmod 2024-09-25created 2024-09-251 edits

Tesla open-sourced the Tesla Transport Protocol over Ethernet (TTPoE) at HotChips 2024

Features

Tesla has joined the Ultra Ethernet Consortium (UEC) to share this protocol and is working to standardize a new high-speed/low-latency fabric for AI/ML/data centers
TTPoE aims to be a non-proprietary, low-cost, decentralized interconnect protocol with distributed congestion control and standard EthernetII frames
Characteristics of TTPoE
- Like TCP, packet loss and retransmission are allowed, but complete delivery is guaranteed
- The initial deployment of TTPoE was in the Tesla Dojo v1 project
  - The protocol runs entirely in hardware and was deployed on a very large multi-exaflops (fp16) supercomputer with tens of thousands or more concurrent endpoints
- The protocol can set up and run links without CPU or OS involvement
The protocol is not complex or clever; it is based on basic principles
- Ethernet transport is essentially moving data from A to B, and should be limited only by physical limits
- In very large systems, centralized congestion management is a foolish attempt; each endpoint should be resilient and self-managing

TTPoE is an interesting attempt to overcome the limitations of the existing TCP protocol in high-performance computing environments
The main goal appears to be minimizing latency and maximizing throughput through hardware offloading and a simplified state machine
TTPoE has the potential to improve data transfer speed and latency in AI and ML
Tesla open-sourcing this protocol will help accelerate innovation in HPC
However, TTPoE seems unlikely to fully replace TCP on general-purpose networks; it is a solution optimized for high-quality dedicated networks. Standardization and ecosystem building will be important for TTPoE to be widely adopted
Protocols with similar functions include RoCE (RDMA over Converged Ethernet) and NVLink

author Zer0 sitemap