By Key Points
- Moore Threads upgrades the KUAE data center for AI, scaling to 10,000 GPUs.
- Large AI models can be targeted by MTLink technology that boosts GPU interconnectivity.
- The company faces export regulations on advanced technology due to US restrictions.
- Strategic partnerships and funding aim to enhance China’s AI capabilities.
Right from time immemorial in the world of AI and high-performance computing (HPC), Nvidia has always been the dominant player.
This is not only because they have cutting-edge GPUs but also because they can scale these processors efficiently across data centers using their own dedicated hardware and software solutions.
However, Chinese GPU maker Moore Threads is stepping up to challenge this dominance with their innovative solutions
Chinese GPU maker Moore Threads’ MTLink fabric tech challenges Nvidia’s NVLink, can now scale to 10,000 GPUs for AI clusters https://t.co/TvDAS1Qlhx
— Anj Bryant (@anjbryant) July 8, 2024
MTLink: Moore Threads’ Answer to Nvidia’s NVLink
Recently Moore Threads has made significant progress in scaling its GPU capacity for AI clusters. As reported by the Science China Morning Post, a single cluster at the company’s KUAE data center now supports up to 10,000 GPUs after upgrading. This has been enabled through their proprietary MTLink technology that links together their MTT S4000 GPUs.
In fact, each of the KUAE servers integrates eight MTT S4000 GPUs with a memory bandwidth of 768 GB/s as well as 128 tensor cores and 48 GB GDDR6 memory respectively.
A cluster containing ten thousand GPUs would involve a staggering number of one million two hundred eighty thousand tensor cores. However, it remains unclear what is possible within such a configuration under real-world conditions since it depends on many factors including software optimization and workload characteristics.
Through strategic moves from financial backers
Regardless of being placed under the US Department of Commerce’s Entity List which restricts access to advanced process technologies; Moore Thread pushes ahead. These conditions present a significant hurdle for a company that was founded in 2020 by a former Nvidia China executive. In terms of raw performance, their GPUs such as the MTT S4000 are currently inferior to those offered by Nvidia. For example, the MTT S4000 is now backdated in comparison with Nvidia’s A100 80 GB GPU which was released in the market last year.
Nevertheless, Moore Threads is not getting stagnant. They have recently gone into strategic partnerships with major state-run telecom operators like China Mobile and China Unicom among others and Gulin Huajue Big Data Technology. Ultimately, these collaborations aim to create new computing clusters for AI development in China.
Financial Backing and Future Prospects
Moore Threads recently closed a huge financing round bringing in about 2.5 billion yuan ($343.7 million approximately). It is this funding that is critical to supporting the company’s ambitious expansion plans and technological advancements however lack of access to advanced manufacturing processes from industry giants TSMC, Intel Foundry or Samsung Foundry remains a big challenge.
But still Moore threads determinedly strive forward with their GPU technology. While they also focus on gaming GPUs alongside AI; their products aren’t yet on par with top graphics cards available on the market.
The Future Path
Moore Threads’ efforts to extend its data center AI capabilities represent a strategic move to take on worldwide leaders such as Nvidia. They are building an infrastructure for more competitiveness in the field of AI and HPC by extending their scale-out capabilities and entering into key partnerships.
However, it is not easy for them to implement their goals smoothly due to technology and geopolitical limitations even though there is an infusion of substantial funding.
It would be interesting to observe how Moore Threads deals with these hurdles as they continue expanding and innovating and what impact this will have on the global GPU market.
