Tencent boosts AI training efficiency without Nvidia's most advanced chips

Tencent Holdings has upgraded its high-performance computing (HPC) network, with an improvement to its artificial intelligence (AI) capabilities, as Chinese tech giants seek ways to boost large language model (LLM) training with existing systems and equipment amid a domestic push for technological self-reliance.

The 2.0 version of Tencent's Intelligent High-Performance Network, known as Xingmai in Chinese, will improve the efficiency of network communications and LLM training by 60 per cent and 20 per cent, respectively, the company's cloud unit said on Monday.

The performance enhancement comes as China looks for ways to advance its AI ambitions with restricted access to advanced chips from Nvidia owing to strict US export rules. China's most valuable tech giant achieved the performance gains by optimising existing facilities rather than trying to compete head-to-head with US rivals such as OpenAI in terms of spending and cutting-edge semiconductors.

Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.

An HPC network connects clusters of powerful graphics processing units (GPUs) to process data and solve problems at extremely high speeds.

Under pre-existing HPC networking technologies, computing clusters were spending too much time communicating with other clusters, leaving a significant portion of GPU capacity idling, according to Tencent. So the company upgraded its network to speed up the communications process while reducing costs, it said.

The Xingmai network can support a single computing cluster with more than 100,000 GPUs, according to the company, doubling the scale from the initial version of the network released in 2023. The improved performance shortens the time needed for identifying problems to just minutes, down from days previously, Tencent said.

Tencent has recently made a big push to strengthen its technologies in the rapidly growing AI field. The Shenzhen-based firm has been promoting its in-house LLMs for enterprise use, and also has services helping other companies to build their own models.

The swift uptake of generative AI service has led China's AI industry to become locked in a price war that has seen all the major players slash their prices to well below what Western firms offer, as companies look for greater commercialisation of their technologies.

In May, Tencent made the lite version of its Hunyuan LLM free of charge while cutting prices for its standard versions. This came after competitors including TikTok owner ByteDance, Baidu and e-commerce giant Alibaba Group Holding made similar moves. Alibaba owns the Post.

The race to improve the efficiency of model training, an energy-intensive and expensive process, is crucial in the price war by making the technology more affordable for the operators and their clients.

Robin Li Yanhong, founder and CEO of Baidu, said in April that the efficiency of training its flagship Ernie LLM had improved more than fivefold within a year. The model's inferencing performance increased 105 times, reducing inferencing costs by 99 per cent, he said.

OpenAI also credited recent efficiency gains as the reason for the much lower pricing of its GPT-4o model launched in May.

This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP's Facebook and Twitter pages. Copyright © 2024 South China Morning Post Publishers Ltd. All rights reserved.

Copyright (c) 2024. South China Morning Post Publishers Ltd. All rights reserved.