en.Wedoany.com Reported - Lenovo held a press conference in Beijing yesterday. Chen Zhenkuan, Vice President of Lenovo and General Manager of China Infrastructure Business Group, announced that by 2027, Lenovo's China Infrastructure Group aims to achieve a target of 100 billion RMB and strives to become the number one in the Chinese server market.

At this press conference, Lenovo launched the Lenovo Wentian Super Node computing power solution. A single node can be equipped with 40 GPUs, delivering FP8 computing power exceeding 28 PFLOPS and HBM memory exceeding 5.76 TB, meeting the training and inference needs of trillion-parameter large models. The total memory access bandwidth exceeds 80 TB/s, with chip-to-chip P2P one-way latency at the hundred-nanosecond level, breaking the communication bottleneck of ten-thousand-card cluster collaboration. A single node supports 40 cards and Scale-out horizontal expansion of clusters, and is also backward compatible with a 32-card configuration, accommodating training, inference, and development testing of various scales. It adopts a cable-free orthogonal direct plug-in architecture, compatible with standard 19-inch chassis, compressing the cluster deployment cycle to just a few hours, significantly lowering the barrier to deploying large-scale computing power clusters.

Lenovo also released the Wanquan Heterogeneous Intelligent Computing Platform V5.0, featuring two major core upgrades. The cluster training and inference acceleration technology achieves comprehensive leadership in large model training and inference performance through core technologies such as a layered decoupled PD architecture and KV Cache shared cache optimization, significantly improving cluster resource utilization. The core model compilation optimization technology enables adaptive matching of computation graphs and automatic generation of operators for different models, deeply adapting to the diverse computing chip ecosystem and enhancing the overall computational efficiency of the training and inference process.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com