The previously announced Nvidia Blackwell B200 accelerators are delayed by several months due to a design flaw.

Nvidia will be forced to delay the launch of its super-powerful Blackwell GB200 GPUs by several months. The Blackwell GB200 is the successor to the Nvidia Hopper accelerator family and should find a home in HPC clusters training new large AI models. The GB200 was announced as by far the most powerful chip of its kind.

The first samples of Blackwell were shipped a few days ago, but now the launch momentum has come to an abrupt halt. According to The Information, TSMC, which makes the chips for Nvidia, has discovered a design flaw in the chip.

Connection error

According to the information, the problem is related to the interconnect processor that connects two Blackwell GPU chiplets on a single Nvidia GB200. New designs must now be tested together with TSMC. The error was discovered exceptionally late, which, according to Nvidia, led to the launch of the Blackwell B200 GPUs being delayed by at least three months.

The company informed Microsoft, Meta, Google and other customers. These technology giants are eagerly awaiting the B200, as this chip has the much-needed additional power to develop new and even more powerful generative AI LLMs. Google alone is said to have an outstanding order for 400,000 affected GB200 chips, which corresponds to a purchase price of more than ten billion euros.

Simpler variant

Nvidia is now said to be planning to produce a simpler version of the Blackwell chip with just a single GPU. Such a chip would still be very powerful, but would not involve the complexity of the connection. In theory, the chip can roll off the production line much faster.

The error will disrupt the plans of major customers somewhat, but is unlikely to have major consequences. Nvidia is the absolute market leader in the AI ​​segment with a market share of 90 percent and more. The Blackwell GB200 also has no real competition. Other chips from AMD, for example, are competitive with today’s market, but theoretically cannot keep up with the planned GB2200.

Not just Nvidia

Modern chips come with a lot of complexity that designers should deal with before launching them. In addition to Nvidia, Intel and AMD are also experiencing problems. For example, Intel is struggling with the degradation of its Raptor Lake (Refresh) chips due to a bug, and AMD has delayed the launch of the Ryzen 9000 to fix an alleged problem.