NVIDIA Blackwell B30A being prepped for China: faster than H20, half the performance of B300

NVIDIA is reportedly cooking up the Blackwell B30A chip that will replace Hopper H20 in China, where the new B30A will beat the H20, but it will offer just half the performance of the Blackwell Ultra B300 AI GPU.

2

In a new report from Reuters, NVIDIA has reportedly finalized the new B30A chip design for China that is compliant with US regulations, with B30A being faster than H20, but much slower than higher-end AI chip offerings like B200 and the upcoming B300.

NVIDIA's purported B30A will feature a single-chip design, compared to the dual chiplets found on B200 and B300. NVIDIA's upcoming Blackwell Ultra B300 features 50% more performance than B200, with 50% more HBM3E memory, and twice the interconnect speeds. The highest-end B300 AI GPU will feature two reticle-sized GPUs with a total of 15 PetaFLOPS of FP4 compute performance, and up to 288GB of HBM3E memory.

The new B30A with half the specs of the B300 means we can expect up to 144GB of HBM3E memory, which is 48GB more of VRAM compared to the 96GB on the H20 that gets sold in China. The H20 has 41% less CUDA cores and 28% less performance than the Hopper H100 AI GPU.

Reuters' sources explained: "The new chip, tentatively known as the B30A, will use a single-die design that is likely to deliver half the raw computing power of the more sophisticated dual-die configuration in NVIDIA's flagship B300 accelerator card".

The outlet continues: "A single-die design is when all the main parts of an integrated circuit are made on one continuous piece of silicon rather than split across multiple dies. The new chip would have high-bandwidth memory and NVIDIA's NVLink technology for fast data transmission between processors, features that are also in the H20 - a chip based on the company's older Hopper architecture".

The successor to Blackwell B30 (and we're guessing the B30A now) is the Rubin R30 AI GPU that's not expected to hit the shores of China until 2028.