Tencent Cloud launches "Model as a Service" (MaaS) for large language models. The service is powered by a new GPU from Nvidia.
Tencent Cloud MaaS offers several built-in language models covering different industries such as finance, media, tourism, or manufacturing, which can be further specialized by companies for their purposes. Alternatively, custom models can be trained via the cloud service.
The company provides a range of tools, including data labeling, training, evaluation, testing, and model deployment tools.
Tencent first in China to use new Nvidia GPUs
The company is working with numerous partners to build a Chinese language model ecosystem and has already deployed more than 50 language model-based industry solutions across more than 10 industries, according to Tang Daosheng, Tencent's senior executive vice president and head of the Cloud and Smart Industries Group.
Similar services are offered by Baidu in China. However, Tencent is the first Chinese company to open a high-performance computing cluster with an Nvidia H800 GPU in April. In addition to the Nvidia GPU, Tencent also relies on its own StarLake servers for its data center.
Export restrictions hit China's AI industry
In China, Nvidia's H800 is the fastest AI accelerator currently available. However, the H800 is a scaled-down version of Nvidia's top-of-the-line H100, which cannot be sold in China due to US export restrictions imposed by the CHIPS Act. The restrictions affect Nvidia's A100 and H100 GPUs, which are used for AI training in most of the world's data centers.
The company has reduced the chip-to-chip data transfer rate of the A100 from 600GBps to 400GBps and is selling the card as the A800. According to Chinese insiders, for the H800 the H100's transfer rate has been halved from 600GBps to 300GBps. In AI training, this can make a big difference, so companies with access to H100 cards could have an advantage over Chinese companies.
In China, Alibaba Group, Baidu, and Tencent have all chosen Nvidia's H800 cards.