The first "Zhenwu" WanKa Intelligent Computing Cluster in the Guangdong-Hong Kong-Macao Greater Bay Area is now online

Mars Finance News—On April 7, the first “Zhenwu” 10,000-card intelligent computing cluster in the Guangdong-Hong Kong-Macao Greater Bay Area, built jointly by China Telecom and Alibaba Cloud, was officially launched at the Shaoguan data center. As the first deployment project in the Greater Bay Area under the national “ultra-large-scale intelligent computing cluster” new infrastructure initiative, this cluster achieves end-to-end independent R&D across the full chain—from chips, to cloud platforms, to model applications—marking that China’s advanced computing power is moving from breakthroughs in high-end performance toward large-scale industrial implementation.

The cluster achieves multiple key breakthroughs at the technical level. Through technological innovations such as high-performance RoCE networking between cards and dual-plane multi-rail communication, the cluster’s end-to-end network latency is as low as 4 microseconds, while network peak utilization exceeds 95%. It can efficiently meet the communication needs of ultra-large data flows such as AllReduce and AlltoAll during large-model training, and it can stably support the pre-training and inference tasks of thousand-billion-parameter-level large models. The Zhenwu chip complete system provides 1.5T of large graphics memory, and the inter-card interconnect bandwidth exceeds 700GB/s, providing solid support for a domestic computing-power foundation. (Wide-angle observation)