NVIDIA prepares B30A Blackwell AI accelerator for China

NVIDIA has created a China-specific variant of its Blackwell Ultra accelerator called the B30A, according to a Reuters report cited in the article. The B30A uses the same base die as the dual-die B300 but embeds a single die on the package to comply with export restrictions. This follows NVIDIA´s earlier export-modified H20 accelerator, which recently received an export license. The article notes that the U.S. administration granted NVIDIA access to an export license that enables these China-targeted products.

Technical expectations for the B30A are drawn directly from the Blackwell Ultra´s dual-die numbers. The B300 in dual-die configuration delivers 15 petaFLOPS at FP4, 7.5 petaFLOPS at FP6/FP8, 3.75 petaFLOPS at FP16/BF16, and 1.88 petaFLOPS at TF32. By embedding a single die, the B30A is expected to halve those figures across the listed precisions. The article emphasizes that NVIDIA will retain mission-critical features such as high-bandwidth memory (HBM) and NVLink on the B30A, which support both training and inference workloads despite the reduced die count.

The report frames the B30A as a response to Chinese Artificial Intelligence labs seeking higher performance than previous export-modified options. The article suggests these chips may be sufficient to compete with current Chinese-designed AI accelerators when combined with NVIDIA´s software stack and performance profile. By offering a single-die variant that preserves key interconnect and memory features while meeting export constraints, NVIDIA is again positioned to supply high-performance accelerators to the Chinese market for domestic needs and laboratory deployments.