Nvidia Reportedly Delays Blackwell Due to 'Design Flaw'

The company has told some partners it's going to need a few more months due to an issue it discovered late in the production process.

In March, Nvidia officially unveiled its next-generation Blackwell data center architecture, saying it would arrive later this year. Now that it's August, it's crunch time for Nvidia, but the company might not meet its deadline. It's now being reported the company is pushing back Blackwell's deployment plans by a few months due to a design flaw it discovered late in the process.

This surprising development comes from The Information, a paywalled news site. However, The Verge notes that the site says it has two sources who worked on the Blackwell project and a source from Microsoft. The sources stated that Nvidia has told Microsoft and at least one other cloud partner that it needs at least three more months than originally planned to provide Blackwell B200 chips to its partners due to the design flaw it has discovered.

 The GB200 is the flagship AI chip and consists of two GPUs connected by a 10TB/s interconnect along with a Grace CPU. . Credit: Nvidia

The delay will likely push the rollout of Blackwell into early 2025 instead of late 2024 as expected. An Nvidia PR person confirmed to The Verge that production of Blackwell should "ramp in 2H" of 2025. It's also being reported that companies like Meta, Microsoft, Google, and other tech monoliths have already placed orders worth tens of billions of dollars with Nvidia for Blackwell chips.

Nvidia's March announcement for Blackwell included the text, "Blackwell-based products will be available from partners starting later this year," but that might no longer be the case. The company is reportedly working with TSMC on a new set of validation tests to resolve the issue. Luckily for Nvidia, the H100 GPU that the Blackwell B200 chip will theoretically replace is still in high demand, but this unplanned discovery could affect Nvidia's Q4 balance sheet.

Source