Analysis of Nvidia’s China‑Specific Cut‑Down GPUs: H20, B20, and B40

Amid escalating U.S. export controls, Nvidia has released a series of China‑specific, performance‑restricted GPUs—starting with the H20, a limited‑capability version of the Hopper‑based H100, followed by the B20 and B40 (6000D) chips built on the newer Blackwell architecture.

The H20 features 96 GB of HBM3 memory, 4.0 TB/s bandwidth, and reduced FP8/FP16 performance (296 TFLOPs and 148 TFLOPs respectively), achieved by cutting 41 % of GPU cores, making it unsuitable for trillion‑parameter model training but still viable for certain inference workloads.

The B20, a cut‑down variant of the upcoming B200, uses a 4 nm process and NVLink interconnects to enable cluster‑scale performance despite lower per‑chip capability; it is expected to adopt GDDR7 memory, differing from H20’s HBM3, and target niche markets such as Huawei’s 910C.

The B40 (6000D) continues the trend with GDDR7 memory, approximately 1.7 TB/s bandwidth, and NVLink speeds around 550 GB/s, while retaining CUDA support. Its price is projected near $7,000, raising cost concerns given the reduced performance.

U.S. restrictions have simultaneously spurred domestic development, with Chinese firms like Huawei (Ascend series) and Cambricon advancing their own AI accelerators, and companies such as Biren (BR100) showcasing chiplet‑based designs that achieve PFLOPS‑level performance.

Overall, the evolution from H20 to B20 and B40 illustrates how export policies reshape the Chinese AI hardware ecosystem, prompting both reliance on cut‑down foreign chips and accelerated investment in indigenous GPU technologies.