Artificial Intelligence 8 min read

Highlights of Chinese Enterprises at the 2024 OCP Global Summit: AI Network Architecture, High‑Performance Cooling, and WAN Innovations

The 2024 OCP Global Summit in San Jose showcased Chinese tech leaders like Alibaba Cloud and ByteDance presenting cutting‑edge AI network architectures, liquid‑cooling solutions, SRv6 deployments, high‑performance data‑center designs, and future WAN routing innovations, underscoring China's growing influence in AI infrastructure worldwide.

Alibaba Cloud Infrastructure
Alibaba Cloud Infrastructure
Alibaba Cloud Infrastructure
Highlights of Chinese Enterprises at the 2024 OCP Global Summit: AI Network Architecture, High‑Performance Cooling, and WAN Innovations

The 2024 Open Compute Project (OCP) Global Summit was held from October 14‑17 in San Jose, California, attracting a record 7,000 participants; Chinese companies, led by Alibaba Cloud, demonstrated strong innovations in AI network architecture, liquid‑cooling, SRv6, and wide‑area networking.

Key Chinese contributors—including Alibaba Cloud, ByteDance, Wiwynn, Micas, and Edgecore—delivered multiple technical talks covering AI‑focused network designs, SRv6, performance optimization, and advanced switch software, while ByteDance highlighted AI training cluster networking advancements.

Alibaba Cloud presented a 51.2 Tbps Ethernet switch cooling case study, proposing four primary solutions: higher‑bandwidth switch chips, longer DAC cables, low‑power LPO optical modules, and CPO (co‑packaged optics) chips, to address the rising heat density of high‑density AI clusters.

Two optimal air‑cooling approaches were explored—environment‑controlled layout optimization and precise thermal simulation—alongside a novel liquid‑cooling design using single‑cold‑plate modules that saved over 800 W of power without significantly raising material costs.

In addressing network stability for Alibaba’s massive compute clusters, engineers described global traffic monitoring, high‑precision flow analysis, and an Alternating DSCP Marking (A.M.D.) scheme that dramatically improved reliability for AI/ML training workloads such as all‑reduce and all‑to‑all operations.

The seventh‑generation High‑Performance Network (HPN 7.0) architecture was unveiled, featuring a "dual‑uplink + multi‑track + dual‑plane" design, 51.2 Tbps single‑chip switches, 400 G NICs, and custom Solar‑RDMA and ACCL libraries, scaling to 100 k cards and delivering a 14.9 % performance gain for large‑model training.

ByteDance engineers introduced a Scheduled Fabric Ethernet architecture that supports thousands of servers with time‑based scheduling and bandwidth allocation, improving throughput and latency while proposing standardization to the broader industry.

Further research on multi‑plane topologies presented optimal path selection techniques, leveraging precise measurements and CPO technology to maximize link utilization and cross‑plane efficiency.

The Phoenix Wing initiative highlighted Alibaba Cloud’s progress in deploying SRv6 via the SONiC platform, showcasing open‑source milestones, collaborations with Cisco, Microsoft, and Inspur, and a vSONiC virtual testbed to accelerate WAN adoption.

A live SONiC demo demonstrated code optimizations that reduced network fault recovery packet‑loss windows from nearly one minute to just 2 ms, markedly enhancing network stability.

Overall, the 2024 OCP summit underscored the prominent role of Chinese enterprises in advancing AI infrastructure and network architecture, positioning them for an increasingly influential presence on the global technology stage.

High Performance ComputingData Centerliquid coolingSRv6AI networkingOCP Summit
Alibaba Cloud Infrastructure
Written by

Alibaba Cloud Infrastructure

For uninterrupted computing services

0 followers
Reader feedback

How this landed with the community

login Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.