Nvidia's B40 GPU Aims to Navigate China Export Restrictions

Nvidia is reportedly preparing to launch a new GPU accelerator aimed at Chinese consumers, which is expected to enter production as early as June. This new product, tentatively named the B40 or RTX Pro 6000D, is designed to circumvent the recent export restrictions imposed by the United States. These restrictions have particularly affected Nvidia's H20 from its latest Hopper generation and AMD's Instinct MI308.

According to sources cited by Reuters, the B40 is similar to the RTX Pro 6000, a professional variant of the GeForce RTX 5090 intended for use in workstations and servers. The US government has linked GPU exports to the memory bandwidth and interconnect speed between multiple GPUs, which are critical for training large artificial intelligence models. The new B40 GPU will reportedly utilize GDDR7 memory instead of the more advanced High-Bandwidth Memory (HBM), which Nvidia is said to have avoided due to the restrictions. The GDDR7 memory is capable of transferring data at a rate of 1.7 to 1.8 terabytes per second, which is significantly lower than the 8 terabytes per second achieved by Nvidia's B200 and B300 models using HBM3(e).

The RTX Pro 6000 and GeForce RTX 5090, featuring the GB202 chip and GDDR7 memory on a 512-bit interface, can also reach the same bandwidth of 1.7 to 1.8 terabytes per second. This configuration allows for up to 96 gigabytes of RAM on a single card. Given the imminent production timeline, it seems unlikely that Nvidia will develop an entirely new chip specifically for the Chinese market, making the use of the existing GB202 graphics chip a practical decision.

However, the absence of Nvidia's NVLink interconnect means that multiple GPUs will need to communicate over the slower PCI Express interface. If the US government decides to enforce speed limitations similar to those established during President Biden's administration, Nvidia may need to throttle the GPU's performance. Previously, the focus had shifted away from performance limitations in the discussions surrounding export restrictions, indicating a potential change in strategy.

Nvidia is reportedly targeting a price range of $6,500 to $8,000 for the B40 or RTX Pro 6000D, significantly lower than the $10,000 to $12,000 price range of the H20. The high-end models, such as the B200 and B300, are priced even higher.

Moreover, Nvidia is said to be in the process of designing a dedicated chip for the Chinese market, which could help reduce production costs by eliminating unnecessary components like 3D rendering capabilities. This move could also lessen the physical footprint of the chip. However, creating a distinct chip for China carries risks, particularly if the US government alters its export regulations and prohibits the sale of such models.