Design Flaws Delay Shipment of NVIDIA's Blackwell Chips

Due to design flaws, the shipment of NVIDIA's Blackwell chips may face a delay of three months or more, which might affect customers such as Meta, Google, and Microsoft. But ***ysts say the potential production delay has a limited impact on the demand for artificial intelligence chips from the chip giant NVIDIA.

According to a report by The Register a few days ago, due to the complexity of TSMC's packaging technology CoWoS, NVIDIA will delay the shipment of Blackwell GPUs to the first quarter of 2025. NVIDIA recently informed Microsoft that the most advanced model in the Blackwell series will be affected by the delayed launch. Customers such as Microsoft and Meta have placed orders worth billions of dollars for new GPUs to drive their respective artificial intelligence services.

A report by the semiconductor research company SemiAnalysis said the main issue behind the delay in GPU shipments is related to the physical design of NVIDIA's Blackwell series. Blackwell is the first high-volume design to use TSMC's CoWoS-L packaging technology.

CoWoS is a method of using interconnected small chips to design more complex and advanced products. These interconnected small chips are typically system-on-chips (SoCs) and one or more high-bandwidth memory (HBM) dies, but the complexity of CoWoS-L is completely different from that of CoWoS-S.

On the other hand, TSMC also doesn't have sufficient CoWoS packaging capacity to meet the demand. SemiAnalysis said that TSMC has built CoWoS-S production capacity in the past few years, mainly serving NVIDIA, but now the GPU manufacturer is shifting its products to CoWoS-L. TSMC is building new fabs for the production of CoWoS-L and urgently needs to convert its old CoWoS-S capacity to keep up with the demand.

But ***ysts say concerns about the delayed launch of NVIDIA's artificial intelligence chips may be exaggerated, and it is expected that this will not have a significant impact on NVIDIA's revenue or demand. According to a report by Reuters, Stacy Rasgon, an ***yst at Wall Street investment bank Bernstein, said that despite recent concerns, it is clear that the demand level continues to rise, and the capital expenditure expectations of hyperscale enterprises are all continuing to grow. If the Blackwell chips are delayed, the sales of NVIDIA's older Grace Hopper chips will fill the gap. "NVIDIA currently has a very large competitive window, and we don't think a three-month delay will lead to a significant change in market share."

A spokesperson for NVIDIA did not deny these reports and said the demand for Hopper chips is very strong, extensive sampling of Blackwell has begun, and production is expected to increase in the second half of the year.

NVIDIA holds more than 80% of the artificial intelligence chip market share. In March this year, NVIDIA launched the Blackwell B200 GPU and GB200 superchip. The training performance of the Blackwell GPU is 4 times that of the previous-generation Hopper GPU, the reasoning performance is 30 times, and the energy efficiency is about 25 times. NVIDIA's CEO Jensen Huang said in May that the latest Blackwell series of artificial intelligence chips will be shipped in the second quarter. NVIDIA's Chief Financial Officer Colette Kress said in May that the Blackwell chips are likely to be in short supply "until next year".