On April 14th, Tencent Cloud released a new generation of High-Performance Computing Cluster aimed at large-scale model training. The overall performance has been improved three times compared to the past.
This cluster uses Tencent Cloud’s self-developed servers, equipped with the latest generation of NVIDIA H800 GPUs. The servers are connected by 3.2T interconnect bandwidth, providing high-performance, high-bandwidth and low-latency cluster computing power for large-scale model training, autonomous driving, scientific computing and other applications.
In October last year, Tencent completed the first trillion-parameter AI model training – the Hunyuan NLP model. The training time was reduced from 50 days to 11 days. If based on the new generation cluster, the training time will be further shortened to 4 days.
Currently, Tencent Cloud’s distributed cloud-native scheduling has a total scale of over 150 million cores and provides 16 EFLOPS of intelligent computing power.
Tencent Cloud’s new generation cluster will serve not only large-scale model training but also be applied in various scenarios, including autonomous driving, scientific computing, and natural language processing.