Tech giant Tencent launched a new high-performance computing cluster (HCC) on Friday in a bid to better train large language models (LLM).
The new HCC uses Tencent's self-developed StarLake server and features Nvidia's latest generation H800 GPU.
According to Tencent's tests, the computing performance of this cluster has improved threefold compared to the previous generation, making it currently the most powerful large-scale model computing cluster in China.
In October of last year, Tencent completed the training of the first LLM with a trillion parameters - HunyuanAide. Using the same dataset, training time was reduced from 50 days to 11 days. If based on the new generation cluster, the training time will be further reduced to 4 days.
Traditional single-server computing power is insufficient when LLMs enter the arena of a trillion parameters. A large-scale computing cluster requires numerous servers connected through an RDMA network.
