Design Flaw Delays Nvidia's Blackwell Chips Until 2025

Nvidia customers who have ordered its new Blackwell flagship line of chips are facing delays as the company grapples with design flaws.

Nvidia discovered a flaw in its new B200 chips “unusually late in the production process,” according to The Information. 

The B200 chips were expected to be in partners' hands later this year, but the delay could impact production targets, forcing some shipments to be pushed back to 2025.

Manufacturer Taiwan Semiconductor Manufacturing Company (TSMC) uncovered the reported design flaw, which affects the processor die connecting two Blackwell GPUs on a single board.

Nvidia will now have to rework the design before conducting further production tests with TSMC before it can move on to mass production.

Companies including Meta, Microsoft and Google reportedly spent “tens of billions of dollars” to get the new chips, which can run massive AI models at 25 times lower costs than previous Nvidia hardware.

Meta wants more hardware to expand its infrastructure, while Microsoft wants to secure the chips for its and OpenAI’s training efforts. Google, meanwhile, is believed to have ordered more than 400,000 Blackwell chips, despite touting the power of its own custom hardware, Tensor Processing Units (TPUs).

To address the delays, Nvidia will extend the life span of its former flagship line of chips, Hopper, according to chip industry analyst firm SemiAnalysis.

Related:Nvidia Unveils AI Tools to Accelerate Humanoid Robot Development

In addition to offering Hopper chips, SemiAnalysis suggests Nvidia will offer a chip that works on the current die system: the B200A.

The initial B200A is designed to power lower to mid-range AI systems before being eventually replaced by the B100 and B200 chips once Nvidia has resolved the design flaw.

Notably, the die used in the B200A will also be featured in the chip Nvidia is specifically designing for the Chinese market, the B20.

News broke last month that Nvidia is working on a chip to recapture the Chinese chip market after U.S. export controls prevented it from selling powerful hardware to clients in the country.

This story first appeared in IoT World Today's sister publication AI Business.