Nvidia is in no hurry to start shipping B20 accelerators to China due to concerns about new US sanctions

Apparently, Nvidia’s willingness to play “regulatory leapfrog” with the American authorities, who regularly increase export restrictions on the supply of computing accelerators to China, has its reasonable limits. B20 accelerators with Blackwell architecture prepared for this country may be delayed on their way to Chinese customers due to fears of new US sanctions.

Image source: NVIDIA

When the US authorities introduced restrictions on the export of computing accelerators of a certain performance level to China in October 2022, Nvidia quickly offered its Chinese customers the A800 and H800 accelerators adapted to these requirements, based on the Ampere and Hopper architectures, respectively. A year later, they also came under a new wave of US sanctions, and then the company offered Chinese customers H20, L20 and L2 accelerators adapted to the new requirements. Wanting to provide Chinese customers with access to the latest Blackwell architecture, it is rumored to be preparing B20 accelerators for announcement, but due to uncertainty about future US export control policies, it is in no hurry to begin shipping them to China.

The South China Morning Post reported this this week, citing its own sources. Nvidia initially planned to begin shipping B20 accelerators to Chinese customers in the fourth quarter of this year, but is now in no hurry to do so due to concerns about the introduction of new sanctions that will cast doubt on the very possibility of such deliveries. Even limited at the level of specific performance of each accelerator, B20 remains attractive to Chinese customers, since it allows efficient scaling of performance by combining multiple accelerators at the cluster level.

The Information publication simultaneously reported that Nvidia is preparing a GB20 accelerator adapted for China, which is based on a truncated GB200. In addition to the GPUs themselves, such accelerators also use Nvidia-developed Grace family central processors based on Arm-compatible architecture.