Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Huawei has started the delivery of its advanced artificial intelligence chip “cluster” to Chinese clients who are increasing orders after being cut off from Nvidia’s semiconductors because of Washington’s export restrictions.
The Shenzhen-based tech conglomerate has sold more than 10 sets of CloudMatrix 384, which links a large sum of chips together, according to two people with knowledge of the matter.
Those who received the first deliveries include data centres that serve Chinese tech companies, one of the people said.
Industry analysts have said they are impressed by the speed at which Huawei has built and begun to ship CloudMatrix, a system that connects 384 AI processors to provide the computing power needed to develop AI models and services.
“The development of Huawei’s CloudMatrix 384 means China now has an AI system capable of beating Nvidia’s,” said Dylan Patel, founder of chip consultancy SemiAnalysis.
Huawei’s new product comes as Chinese customers have been further restricted from Nvidia’s AI chips. The US group said this month it would take a $5.5bn earnings hit after President Donald Trump made its H20 chip, tailored to comply with earlier export controls, require a special licence to be sold to Chinese customers.
Huawei is accelerating its chip development to help domestic tech companies compete with their global peers despite their restricted access to the world’s best technology in the area.
It has told local clients its CloudMatrix performs significantly better than Nvidia’s NVL72, a popular AI cluster used by US tech giants, which consists of 72 of its GB200 chips, in terms of computing power and memory, according to a company presentation reviewed by the Financial Times and people with knowledge of the matter.
CloudMatrix 384 uses Huawei’s Ascend 910C chips, which by themselves underperform Nvidia’s GB200 processors.
But Huawei has used a larger quantity of the chips — connected by its “super node”, through which all processors in the set are connected using optical technology — to boost the cluster’s overall performance.
In a company presentation, Huawei claimed CloudMatrix outperformed NVL72 in key metrics, saying it was 67 per cent higher in compute and more than three times larger in aggregate memory capacity.
“It is compensating for weaker individual chip performance with advanced networking to boost the chip performance in a cluster,” said Patel at SemiAnalysis.
Industry analysts said Huawei had leveraged its background in telecommunications to improve its chip systems’ performance.
Despite the strong metrics, CloudMatrix 384 has several disadvantages compared with Nvidia’s rival product, said industry experts.
Due to the use of more chips, CloudMatrix 384 has much higher energy consumption, leading to larger electricity bills. Huawei’s software systems, compared with Nvidia’s Cuda, require more maintenance from experienced engineers, resulting in operational manpower costs three to five times larger.
Still, given China’s abundance of power and engineers, CloudMatrix was a viable solution for clients, especially when they were cut off from Nvidia’s most advanced technology, said people with knowledge of the sales.
CloudMatrix 384 sells for about Rmb60mn ($8.2mn) a set, and the exact price depends on the individual contract, according to industry insiders. This compares with about $3mn for Nvidia’s NVL72, according to analyst estimates.
Nvidia said it had range of prices depending on original equipment manufacturer and customer specifications.
Huawei declined to comment.
Read the full article here
