Skip to content

Illegally bypassing export restrictions, Nvidia allegedly planning to release a scaled-down B300 model exclusively for the Chinese market.

facing the possible need to replace aged H200 NVL PCIe cards or opting for an alternative solution instead

Undermining export restrictions, Nvidia allegedly planning to release a scaled-down version of the...
Undermining export restrictions, Nvidia allegedly planning to release a scaled-down version of the B300 model specifically for the Chinese market

Illegally bypassing export restrictions, Nvidia allegedly planning to release a scaled-down B300 model exclusively for the Chinese market.

Nvidia has announced the development of a new AI accelerator, the B30A GPU, specifically designed for the Chinese market. Compliant with US export restrictions, the B30A aims to offer improved performance over existing models while adhering to regulatory limits.

Key Details and Specifications of the Nvidia B30A:

The B30A is based on Nvidia's Blackwell architecture, the same generation as the newest consumer RTX GPUs. It features a single-chip (single-die) design, unlike higher-end models such as the B200 and B300, which employ a dual-chiplet design.

In terms of performance, the B30A is expected to deliver approximately 7.5 PetaFLOPs FP4 compute performance, which is about half the compute power of the B300. However, it boasts up to 144GB of HBM3E memory, a substantial increase over the 96GB HBM3 on the Hopper H20 model previously sold in China.

The B30A retains Nvidia’s NVLink technology for high-speed chip-to-chip communication and uses TSMC’s CoWoS-S packaging, accommodating one compute chiplet and four HBM3E stacks, optimizing manufacturing costs and complexity.

Intent Behind the B30A:

The B30A is designed to comply with US export restrictions aimed at limiting the performance of advanced AI chips sold into China. It offers improved capability over the H20 but caps performance roughly at 50% of the leading B300 chip to meet regulatory requirements.

Nvidia’s efforts coincide with a reported agreement involving remittance of 15% of China sales revenue to the US, reflecting geopolitical export control measures. The B30A's design simplifications and performance caps seek to balance competitive performance for Chinese customers while adhering to US government export controls.

Potential Impact and Implications:

The B30A's release may impact the demand for Nvidia accelerators in China, as Beijing pressures cloud providers and model developers not to employ Nvidia kit, particularly for sensitive government workloads.

However, the B30A could be a viable replacement for the aging H200 NVL due to its power consumption and form factor. Moreover, the B30A may not be limited to China due to potential demand from customers outside of China.

The B30A, if offered, could reflect a hedging of bets should US trade policy shift, making it attractive to customers beyond China. President Trump suggested he was open to the idea of allowing sales of Blackwell-based chips to China if the performance were "enhanced in a negative way" by 30-50%.

[1] Nvidia's B30A GPU for the Chinese market: A power-efficient AI accelerator [2] Nvidia's B30A chip: A cut-down version of the RTX Pro 6000 server chip [3] Nvidia's B30A chip: A potential replacement for the aging H200 NVL [4] Nvidia's B30A chip: Balancing performance for the Chinese market [5] Nvidia's B30A chip: A look at the new AI accelerator for China

Read also:

Latest