NVIDIA’s H20 AI GPU: Set for Mass Production in China by Q2 2024, Fully Compliant with US Policies

NVIDIA, the US-based technology giant, has announced a delay in the launch of its new artificial intelligence HGX H20 AI chip, in China until the first quarter of 2024. This chip is the most powerful of the three China-specific chips the company has ever developed to comply with the new US export regulations, wish it was the same for the RTX 4090D.

Source: MyDrivers

The H20 chip was expected to hit the market in November 2023 but faced some technical challenges with server manufacturers who were supposed to integrate the chip into their systems. The chip is part of NVIDIA’s strategy to maintain its market share in China, one of the world’s largest and fastest-growing markets for AI chips.

The U.S. government has imposed strict export restrictions on advanced AI products that NVIDIA and other U.S. companies can ship to China, citing national security concerns. To circumvent these restrictions, NVIDIA designed three new chips – the H20, the L20, and the L2 – with lower computing capabilities than its flagship products, the A800 and the H800.

Nvidia H20 Ai Chip – Technical Specifications

The new chips use a different GPU, the AD102, which has lower memory sizes of 48GB and 24GB GDDR6. They can only deliver a maximum compute performance of 4800 TOPS over the PCIe interface, much lower than the 10,000 TOPS that the A800 and the H800 offer. Despite these limitations, the new chips still include some of the latest NVIDIA AI features, such as the Tensor Core, the Multi-Instance GPU, and the NVLink.

The fully unlocked version of the H20 chip also comes with 96GB of HBM3 memory and a memory bandwidth of 4.0 TB/s.

However, NVIDIA is not alone in the Chinese AI market. Its rivals, such as Huawei and AMD, are also vying for a slice of the lucrative pie. Huawei, the Chinese telecom giant, has developed its own AI chips, such as the Ascend series, which have been gaining popularity among Chinese customers. AMD, another US-based technology company, is also planning to launch its AI chip for the Chinese market, the MI300 accelerator.

Summing up

• GPU: Ada-Lovelace / Hopper
• Memory Sizes: 96GB HBM3, 48GB and 24GB GDDR6
• Maximum Compute Performance: 4800 TOPS over the PCIe interface
• AI Features: Tensor Core, Multi-Instance GPU, and NVLink

Conclusion

The delay in the launch of Ai chip indeed underscores the complexities that technology companies like NVIDIA face in navigating regulatory landscapes. The H20 chip is part of its strategic move to comply with new U.S. export regulations. This reflects the company’s efforts to maintain a foothold in the Chinese market after facing restrictions on certain product shipments.

The new chip is the most powerful of the three China-specific chips the company has developed to comply with new US export regulations. It was expected to hit the market in November 2023 but faced some technical challenges with the server manufacturers who were supposed to integrate the H20 chip into their systems. This has resulted in a 1.9% drop in NVIDIA’s shares, impacting the market on a shortened U.S. trading day.

Despite these limitations, the new chips still include some of the latest AI features, such as the Tensor Core, the Multi-Instance GPU, and the NVLink. NVIDIA hopes these features will attract Chinese customers who need high-performance AI solutions for applications such as cloud computing, data analytics, and machine learning.