The Blackwell B200 GPU would be delayed at least three months after an “unusually late” discovery of a flaw during the manufacturing process.
Sources report that to The Information. The design flaw concerns the processor die, which connects two Blackwell GPUs on a single Nvidia GB200 Superchip. This superchip is planned to be the first product to actually use the new B200. The B200 is expected to realize a performance boost and reduce power consumption.
TSMC, which widely produces the chips for Nvidia, is said to have discovered the design flaw. The sources speak of an unusually late discovery. It prompted Nvidia to go back to the drawing board and revise the design of the processor die. After the design flaw is rectified, production tests must still be conducted with TSMC before mass production can take place. This extra time can prevent a defective product from appearing on the market.
Orders
Many major tech companies are looking forward to the new Nvidia chip, which promises to be another step forward as the basis for AI workloads. The Information knows that Microsoft, which is said to have a significant order outstanding, has already been notified about the delay. Other cloud providers have also been notified.
Just a few months ago, Nvidia communicated the end of 2024 as the date for the Blackwell chips. The delay would now push the launch of the new server chips to at least early 2025.
Alternatively, sources claim that Nvidia considers a single-GPU version of the B200. That way, a select number of customer orders can still be filled. Besides Microsoft, which reportedly takes 55,000 to 65,000 GB200 chips, there are tens of billions more orders on the books. Google and Meta also want to purchase the new server chip.
Tip: Nvidia brings less advanced version of B200 Blackwell chip to China