Tech war: Intel and Nvidia continue to push purpose-built chips for training AI systems in China amid US export restrictions

Intel said its Gaudi2 processor, which is not subject to US restrictions, is the company’s answer to Nvidia’s A100 GPU, used for training AI systems
It follows efforts by Nvidia earlier this year to push modified versions of the firm’s US-restricted A100 and H100 GPUs in China

Semiconductors

Ben Jiangin Beijing

Published: 8:00pm, 12 Jul 2023

Why you can trust SCMP

Semiconductor giant Intel Corp has brought its latest processor for artificial intelligence (AI) deep-learning applications to mainland China, where massive demand for US-restricted advanced chips has created a major under-the-counter trade for smuggled graphics processing units (GPUs) from Nvidia Corp.

At a press conference in Beijing on Tuesday, Intel executives presented the company’s Gaudi2 processor, a device that is not subject to US export restrictions, as its answer to Nvidia’s premium A100 GPU, which is widely used for training AI systems.

The latest initiative by Intel – which generated 27 per cent of its total 2022 revenue in China, according to its latest annual report – underscores the continued importance of the vast mainland market to US semiconductor technology providers, despite Washington’s export controls.

It follows efforts by Nvidia earlier this year to push modified versions of its flagship A100 and H100 GPUs in China to comply with US restrictions and keep the supply lines open for its customers on the mainland, where new AI development projects have flourished to create services similar to OpenAI’s ChatGPT.

The Habana Gaudi2 Mezzanine Card was initially launched by Intel in the US on May 10, 2022. Photo: Handout

In August last year, the US Department of Commerce imposed a ban on Nvidia and Advanced Micro Devices (AMD) from selling some advanced chips to China. Nvidia was instructed to stop selling its A100 and H100 GPUs, while AMD was prevented from exporting its MI250 chips to the mainland.

That left China’s AI sector scrambling to buy limited supplies of advanced chips from Nvidia, which has a near monopoly on GPUs used to train AI systems.

The strong demand has even created a fast-growing market for smuggled GPUs, such as the A100 and H100 devices from Nvidia.

The A100 Tensor Core GPU, according to the Nvidia website, powers the world’s leading data centres used for AI, data analytics and high-performance computing applications.

Nvidia Corp’s A100 Tensor Core graphics processing unit. Photo: Handout

For Intel, the demand in China strengthens its commitment to provide clients “with a wide range of hardware selections”, company executive vice-president Sandra Rivera, who also serves as general manager of the firm’s data centre and AI group, said at the press conference in Beijing.

She said the Gaudi2 was designed to lower the barrier to entry and enhance its mainland clients’ ability “to deploy AI through cloud and smart-edge technologies, helping build up China’s AI future”.

Intel announced that it is working with Inspur Group – the world’s second-largest AI server manufacturer, based in eastern Shandong province – to build new Gaudi2-powered machines for the mainland market.

Habana Labs, Intel’s data centre team focused on AI deep learning processor technologies, initially launched Gaudi2 last May in the US, where it said the processor’s training throughput performance was twice that achieved by Nvidia’s 80-gigabyte A100 GPU for the ResNet-50 computer vision model and the BERT natural language processing model.

Chinese firm UCloud raises market uncertainty in securing advanced Nvidia chips

Intel’s latest offering comes at a time when more difficulties potentially loom for China’s AI ambitions.

The US government is said to be considering restricting Chinese companies’ access to American cloud computing services, which would stop Amazon Web Services and Microsoft Corp from using the power of advanced AI chips to benefit their mainland clients, according to a report by The Wall Street Journal earlier this month.

The US was said to be also considering an escalated move to include Nvidia’s remodelled A800 GPU in its export ban, according to a report by The Wall Street Journal last month.