HyperAI
Back to Headlines

Nvidia Set to Launch Blackwell-Based B40 GPUs in China to Replace Banned H20 Accelerators

il y a 6 heures

In the wake of the U.S. export ban on its Hopper H20 accelerators to China, Nvidia is reportedly preparing to launch new Blackwell-based solutions this year, according to Reuters. Mass production of these new GPUs is expected by June, making them widely available in the Chinese market by the third or fourth quarter. Although technical details are still limited, some key information has emerged. The ban on Hopper-family GPUs has created a significant challenge for Nvidia, turning into a regulatory cat-and-mouse game. Before they were officially released, the flagship H100 and H200 accelerators were already subject to export restrictions. In response, Nvidia introduced the H800, which also fell under similar constraints in October 2023. The company then relied on the cut-down H20 as its primary AI solution for the Chinese market until it was recently banned, leading to a $5.5 billion write-off in GPU supply. The upcoming replacement, based on the Blackwell architecture, is tentatively named the RTX Pro 6000D or B40. This GPU is classified as a server-class model and uses traditional GDDR7 memory instead of high-bandwidth memory (HBM). The decision to avoid TSMC's CoWoS packaging technology suggests that the B40 will have a monolithic design rather than a multi-chip module. There are two potential silicon options for the B40: the datacenter-grade GB1XX Blackwell or the consumer-grade GB2XX Blackwell. Given that GB1XX features only HBM controllers at the silicon level, it is unlikely to be the basis for the B40. Instead, the B40 is more likely to be a derivative of the GB202 chip, found in the RTX 5090 and RTX Pro Blackwell 6000, and will lack NVLink support. The estimated price range for the B40 is between $6,500 and $8,000, making it more affordable than the H20 and comparable to Nvidia's global RTX Pro 6000 workstation models. While the HGX H20 could be configured in an 8-GPU setup using NVLink for efficient multi-GPU communication, the B40’s lack of NVLink may present challenges in such configurations. However, Nvidia's latest RTX Pro Blackwell servers manage GPU-to-GPU communication through ConnectX-8 SuperNICs with integrated PCIe 6.0 switches. It is speculated that the B40 might adopt a similar approach, utilizing up to eight GPUs in a server. For scaling beyond this limit, Nvidia’s Spectrum-X networking platform is expected to play a crucial role. These details are preliminary and should be taken with a grain of salt until more information becomes available. To stay updated on the latest developments in this area, follow Tom's Hardware on Google News and make sure to click the Follow button.

Related Links