Chinese GPU start-up Moore Threads upgrades its AI data centre solution despite US restrictions
- The upgrade is the latest example of Chinese firms trying to catch up in AI capabilities despite US export restrictions on advanced chips being shipped to China.
Chinese graphics processor unit (GPU) start-up Moore Threads has unveiled an upgrade to the capability of its artificial intelligence (AI) data centre solution KUAE, boosting the number of GPUs connectable in a single computing cluster to 10,000, as US sanctions continue to hobble its technology progress.
The company, founded in 2020 by a former Nvidia China veteran, said in a statement on Wednesday that the upgrade puts more GPUs into the single computing cluster via high-speed connectivity to provide infrastructure for China’s innovation in AI.
Separately, Moore Threads completed a new round of financing of up to 2.5 billion yuan (US$343.7 million), Chinese news outlet Jiemian reported.
KUAE is a data centre server which incorporates eight Moore Threads MTT S4000 GPUs designed for large language models (LLMs), according to its website. The MTT S4000, based on its in-house developed MUSA architecture, contains 128 tensors cores and 48GB of GPU memory, with memory bandwidth of 768 gigabytes per second. The GPUs can be interlinked by a technology called MTLink, that Moore Threads developed, allowing multiple GPUs to be connected together to form a more powerful system.
The upgrade is the latest example of Chinese firms trying to catch up in AI capabilities despite US export restrictions on advanced AI chips being shipped to China, over national security concerns.
Despite being a leader in China’s booming GPU sector, Moore Threads, which was added to a US trade blacklist in October 2023, still lags Nvidia in terms of product performance, with the US firm temporarily surpassing Microsoft and Apple to become the world’s most valuable company in June with an eye-watering market capitalisation of US$3.34 trillion.
An entry-level version of Nvidia’s A100 GPU, equipped with 40GB of high bandwidth GPU memory and a memory bandwidth of 1,555 GB/s, provides a computing power equal to 312 teraflops, more than three times that of the MTT S4000, according to specifications published by Nvidia.
On Wednesday, Moore Threads signed deals with state-run telecoms operators China Mobile, China Unicom and China Energy Engineering Corporation, and Gulin Huajue Big Data Technology to build three computing cluster projects.
