AI News

---

The new Nvidia Blackwell chip designed for China might perform better than the H20 model.

Nvidia is developing a new AI processor for China that would be more powerful than the H20 model it already distributes there, said individuals who know the plans, according to Reuters. The processor will be built on the firm’s latest Blackwell design.

Last week, US President Donald Trump hinted that higher-end Nvidia chips could be sold in China eventually. Approval is, however, far from certain, as US officials are cautious about allowing Beijing too much access to US AI technology.

The new chip, internally called the B30A, is likely to employ a single-die design. That is, all the essential components of the chip are implemented on one piece of silicon, rather than divided into two dies as in Nvidia’s high-end B300. Individuals with knowledge of the details said this would probably impart around half the power of the B300 but keep the B30A more powerful than the H20.

Similar to the H20, the B30A will come with high-bandwidth memory and NVLink, Nvidia’s solution for transferring data at high speed between processors. While final specs aren’t yet finalized, the company is looking to ship samples to Chinese customers for testing as soon as next month.

Nvidia, in its China strategy statement, stated: “We consider a range of products for our roadmap, so we can be ready to compete to the degree governments permit. All that we supply is with the express consent of relevant authorities and intended only for positive commercial application.”

US Commerce Secretary Howard Lutnick explained to CNBC that Nvidia CEO Jensen Huang has been lobbying for approval. “Of course he would like to sell a new chip to China,” Lutnick added, noting Huang tends to make the pitch directly to the president. “The president hears from our great technology companies, and he’ll determine how he wants to play it. But the fact Jensen is selling a new chip shouldn’t be surprising to anyone.”

Disagreements between the US and China are increasing because of issues related to Nvidia chips.

The issue of how much access Chinese has to high-end AI chips is now one of the key issues of contention between Washington and Beijing. China generated 13% of Nvidia’s sales last year.

Nvidia only resumed sales of the H20 in July, having had them cut short by US regulators in April. The H20 was created in 2023 to comply with export regulations that curbed chip shipments to China.

Donald Trump recently stated that he would consider letting Nvidia sell a reduced version of its new-generation chip in China. Under a wider agreement, Nvidia and its competitor AMD agreed to pay the US government 15% of the revenue of some chip sales to China. Trump also referred to the H20 as “obsolete,” implying that a new chip aimed at China alone would provide “30% to 50% off” [sic] the processing of the flagship model.

Bipartisan Washington lawmakers claim even more diluted versions of AI chips would still provide China with a competitive edge in key sectors. Nvidia and others have replied that if they cut sales to China, customers will switch to domestic vendors such as Huawei. The latter’s latest chip models are reported to equal Nvidia’s in sheer computing capability, though analysts add Huawei is still behind in software performance and memory speed.

China’s state media jumped in with additional pressure by stating that Nvidia’s chips may pose security threats, with regulators warning Chinese companies against purchasing the H20. Nvidia dismissed the suggestion that its hardware was a threat.

A new chip is currently being developed.

In addition to the B30A, Nvidia is also developing another product solely for the Chinese market based on the Blackwell architecture but designed for AI inference operations. Sources state that the RTX6000D processor will be less expensive than the H20 with easier design and reduced specifications.

Reuters earlier reported that the RTX6000D was created to operate just below levels established by US export restrictions. It features regular GDDR memory and operates at 1,398 gigabytes per second – below the 1.4 terabyte-per-second limit established in new restrictions issued in April.

Small supplies of the RTX6000D will arrive in Chinese customers in September.

(Photo by BoliviaInteligente)