Report: Design Flaw Discovery Could Delay Launch of Nvidia’s New Blackwell GPUs

Chipmaker Nvidia Corp. informed Microsoft Corp., Meta Platforms Inc., Google LLC and other customers that the launch of its upcoming, highly anticipated “Blackwell” B200 graphics processing units (GPUs) designed for artificial intelligence tasks will likely be delayed by at least three months.

A report from The Information on Friday evening said that one of Nvidia’s key manufacturing partners had discovered a “design flaw” that would prevent the chips from being introduced on time. The flaw was discovered “extremely late” in the manufacturing process, the anonymous sources said.

The Blackwell GPUs are the successors to Nvidia’s wildly popular and coveted H100 GPUs, which are currently powering the majority of the world’s generative AI applications. They were announced in March and are said to deliver up to 30x performance improvements over the H100, while reducing power consumption by up to 25% for some workloads.

The delay was unexpected, as Nvidia had said just three months ago that its “Blackwell-based products” would be available from its partners this year. They were to be the first major release in a series of AI chip updates from semiconductor companies, with rivals like Advanced Micro Devices Inc. also set to launch their own products in the coming months.

According to The Information, the issue is with the processor chip that combines two Blackwell GPUs onto a single Nvidia GB200 Superchip (pictured), which is the first product to feature the new GPUs. The issue was reportedly identified by Taiwan Semiconductor Manufacturing Co., which mass-produces the chips on Nvidia’s behalf. The discovery apparently forced Nvidia to change its chip design, and it will need several months of production testing with TSMC before it can begin mass-producing the chips as planned.

If true, the delay is likely to cause significant disruption for Nvidia customers, many of whom are planning to launch new AI data centers early next year. Google, for example, is believed to have placed orders for more than 400,000 GB200 chips as part of a deal valued at more than $10 billion. Meta has also placed a similar order, while Microsoft is aiming to have between 55,000 and 65,000 GB200 chips ready for OpenAI by the first quarter of next year.

To compensate for the delay, Nvidia is considering producing a single-GPU version of the Blackwell chip, which would allow it to fulfill initial orders.

It makes sense for Nvidia to delay the launch of its Blackwell GPUs rather than ship potentially flawed products. Such a move is not unprecedented in the industry. Just last month, AMD announced it would delay the launch of its Ryzen 9000 processors due to a simple typo. A delay is better than a catastrophic failure — like what apparently happened to Intel Corp.’s 13th- and 14th-generation Core processors.

The likely cost of Nvidia’s Blackwell GPUs makes it even more important for the company to get the product right. Blackwell’s GB200 Grace AI superchips are said to cost up to $70,000 each, while a complete server rack costs more than $3 million. Nvidia reportedly expects to sell between 60,000 and 70,000 complete servers, so any flaws like those found in Intel’s chips would be hugely expensive and damaging to the company’s reputation.

In any case, Nvidia can probably afford to delay. The company is the undisputed leader in the AI ​​chip market, with many analysts saying it controls as much as 90% of global AI chip sales.

While competitors like AMD and Intel have developed their own AI chips, they have yet to gain significant market traction.

Constellation Research Inc.’s Holger Mueller said it’s almost inevitable that the AI ​​industry will face some delays because the frantic pace of development outlined by Nvidia CEO Jensen Huang means there’s no buffer in case of any production issues like the one that’s apparently happening now.

“For cloud service providers, this will cause headaches in the form of delayed revenues,” the analyst said.

But instead of cloud companies buying GB200 chips, it’s AI companies that will suffer the most because they are the end users, Mueller said. He stressed that many of them were counting on Blackwell’s chips to be available this year when they were planning their product roadmaps and growth strategies. The analyst believes Blackwell potentially won’t launch until early 2025, throwing a lot of plans into disarray.

“Overall, the pace of generative AI development may plateau for a while, but that could be good news for enterprises because it means they have a bit of a break and an opportunity to catch up and think about where the technology can really make a difference,” Mueller said. “The real winner could be Google, which runs all of its native AI workloads on its own TPUs. That could give Gemini a chance to get ahead.”

In response to The Information’s report, an Nvidia spokesperson said: “Hopper has been in very strong demand, with Blackwell already in wide distribution and production expected to begin in the second half of the year.”

The delay is the second setback to hit Nvidia in recent days. Last week, it was revealed that the chipmaker was the subject of two U.S. Justice Department investigations into its AI practices. One is looking into whether Nvidia’s $700 million acquisition of Israeli AI startup Run:AI Inc. may have violated antitrust laws, while the other is investigating claims that the company unfairly pressured cloud computing companies to buy its chips.

Image: Nvidia