Ant Group CTO Calls for Green, High‑Quality Computing Infrastructure to Power AI and Large Models

On November 9, 2023, during the "Digital‑Green Collaborative Transformation" sub‑forum of the World Internet Conference Wuzhen Summit, Ant Group Chief Technology Officer He Zhengyu emphasized that the explosion of AI and large‑model applications is driving unprecedented demand for intelligent computing power, and that the mission of digital‑green co‑development requires the industry to rapidly construct green, high‑quality computing infrastructure.

He called on the industry to strengthen soft‑hardware collaboration and ecosystem building to jointly improve the scale, availability, and tolerance of computing infrastructure.

Ant Group has built a "Wan‑Card" heterogeneous cluster, leveraging innovative technologies to create a green and efficient computing architecture that delivers roughly double the AI inference performance compared with typical industry solutions.

The Chinese government has placed strong emphasis on digital‑green collaborative development, launching pilot projects in ten regions across Hebei and Zhejiang to explore replicable and scalable dual‑transformation experiences.

He explained that in the era of large‑language models and generative AI, the depth and breadth of industry intelligence are expanding, leading to a surge in demand for intelligent computing resources; therefore, improving the efficiency and greenness of compute infrastructure is inevitable.

Ant's green computing technology system comprises four core components: multi‑cloud, multi‑chip heterogeneous hardware integration management; offline‑mixed deployment with automated management; cloud‑native containerization with time‑slice scheduling; and continuous performance monitoring with AI‑driven elastic capacity. This synergy of hardware, algorithms, and engineering has boosted CPU efficiency to 33% in 2022—tripling compute efficiency over three years—and earned the "2022 Digital Technology Enterprise Dual‑Transformation Typical Case" award.

Last week, Ant's "Bailing" large model completed registration, and the green computing system is already providing green compute power for the model. The Wan‑Card heterogeneous cluster achieves hardware efficiency (HFU) above 60%, with over 90% of the cluster’s effective training time, delivering a 3.59‑fold increase in RLHF training throughput and about a 2‑fold improvement in inference performance versus industry benchmarks.

He concluded by urging the industry to jointly cultivate high‑quality green hardware and software R&D, promote standardized green computing platform specifications, interfaces, and evaluation metrics, and accelerate open‑source technology adoption to build a sustainable, high‑quality green compute foundation.