Welcome Here  Shenzhen Mingjiada Electronics Co., Ltd.

Aug. 5 (Bloomberg) -- NVIDIA's next-generation AI GPUs based on the Blackwell architecture have been delayed due to "design flaws," The Information reported, citing people familiar with the matter.

According to the report, NVIDIA Blackwell's launch could be delayed by three months or more, which could affect big NVIDIA customers like Meta, Google and Microsoft.

A Microsoft employee and a person familiar with the matter said that NVIDIA has informed Microsoft of the delay, which also affects the most advanced AI chip in the Blackwell family (B200).

NVIDIA officially released AI chips based on the Blackwell architecture in March this year including the B100, B200, and GB200. the most powerful of these, the GB200 chip, contains two connected Blackwell GPUs and a Grace CPU.

However, while TSMC engineers were preparing for mass production in recent weeks, a design flaw was found on the bare die connecting the two Blackwell GPUs. This defect could lead to a reduction in chip yield or production, and the usual practice is to stop mass production.

Two TSMC employees quoted by The Information noted that the news is unusual for the foundry leader to pause the production line and go back to the design stage so close to volume production.

With unusual design flaws popping up in the later stages of manufacturing, NVIDIA is conducting a new batch of trial production with partner TSMC to clear up trouble spots, the sources said.

TSMC had originally planned to mass-produce Blackwell in the third quarter of this year and deliver it in bulk to NVIDIA in the fourth quarter. Now one of the B200 chips will likely be delayed to mass production in the fourth quarter, and if no other issues arise, shipments will be pushed back to the first quarter of 2025 or later. It's unclear if that will affect the timing of other Blackwell GPUs as well.

Industry insiders say it usually takes three months for cloud service providers to get a large number of clusters deployed and up and running after receiving the chips. That means that if NVIDIA's B200 chip is delayed until the first quarter of 2025, then the real commercial go-live may have to wait until the second quarter of 2025.

A spokesperson for NVIDIA wouldn't comment on a statement it sent to customers about the shipment delay, but told The Information that "production volumes are expected to increase" later this year.

Microsoft, Google, Amazon Web Services (AWS) and Meta declined to comment on the news, while a TSMC spokesperson did not respond to a request for comment.

In other news, trial production of Blackwell GPUs has begun at the end of the second quarter of 2024, and any technical issues related to the original design can still be resolved through the software system. NVIDIA hopes to further improve Blackwell's stability by replacing some photomasks, or "redesigning" it. The redesigned Blackwell has already been completed at TSMC, and it is possible that subsequent delays could be shortened by TSMC catching up, so volume production in Q4 2024 is still possible.