New details have emerged regarding NVIDIA's custom Blackwell-based AI chip, potentially codenamed "B30A," which is being developed specifically for the Chinese market. Earlier reports indicated that NVIDIA had submitted this new product to the U.S. government for an export license, targeting a peak performance level that is 80% of a standard Blackwell GPU. NVIDIA is reportedly planning to provide samples to its Chinese customers for testing this September, with an expected price roughly double that of the current H20 accelerator.
Now, a report from Wccftech sheds more light on the chip's specifications. The B30A is said to be equipped with 144GB of 8-Hi HBM3E memory. Like the H20 it is meant to succeed, it will also reportedly support a 900 GB/s NVLink interconnect. Given the significant architectural improvements of Blackwell over the Hopper generation, a substantial performance increase is anticipated despite the necessary cutbacks.
Initial rumors suggested the B30A would be a single-die design, derived from the dual-die B300 "Blackwell Ultra." This theory posited that both chips share the same base die, but the B30A would use a single-module package to comply with U.S. export control policies. Performance estimates for this configuration are:
FP4: 7.5 PetaFLOPS
FP6/FP8: 3.75 PetaFLOPS
FP16/BF16: 1.875 PetaFLOPS
However, contradicting earlier speculation, the new report suggests the B30A may in fact remain a dual-die chip, similar to the full-fat B300, but with its specifications effectively halved to meet compliance requirements.
Reports indicate that NVIDIA is targeting the fourth quarter of 2025 for the B30A's release. This ambitious timeline means the company will need to secure approval from the U.S. government within the next few weeks. According to previous reports, this arrangement may come with a significant stipulation: NVIDIA could be required to forfeit 15% of its sales revenue from these restricted GPUs in China to the U.S. government.
0 Comments