NVIDIA’s Blackwell AI GPU Family Faces Production Delays Due to Design and Packaging Issues

NVIDIA’s highly anticipated Blackwell AI GPU family is facing major production hurdles, according to a detailed analysis by SemiAnalysis. The issues stem from both the Blackwell GPU architecture itself and the advanced packaging technology employed by TSMC, known as CoWoS-L.

One of the key problems, as reported by SemiAnalysis, is a potential redesign requirement for one of the top routing metal layers and the bump on the Blackwell die. Another issue involves the bridge dies within the CoWoS-L packaging, which may also necessitate a redesign. Furthermore, SemiAnalysis highlights a warping issue in the CoWoS-L packaging caused by heat. This complexity in TSMC’s new CoWoS-L packaging is attributed to its more sophisticated nature compared to the widely adopted CoWoS-S packaging.

NVIDIA’s Blackwell AI GPUs are poised to be the first mass-market chips to utilize TSMC’s cutting-edge CoWoS-L packaging. The company had initially planned for an aggressive production ramp-up, aiming to manufacture over a million chips per quarter. However, the multifaceted issues have disrupted these ambitions.

To mitigate the impact of these delays, NVIDIA is reportedly extending the production lifespan of its Hopper AI GPU series. Additionally, the company is introducing a new variant of the Blackwell family, dubbed the B200A. This new GPU will be packaged using TSMC’s CoWoS-S technology and alternative packaging solutions from other suppliers such as Amkor, ASE, and Samsung. The B200A AI GPU will cater to lower-end and mid-range AI systems, serving as a replacement for the B100 and B200 chips designed for the HGX 8-GPU supercomputing platform form factors.

TSMC, in response to the escalating demand for CoWoS-L packaging, has been actively expanding its production capacity. This includes the construction of a new packaging plant named “AP6” and the conversion of existing CoWoS-S capacity at AP3 to CoWoS-L capacity.

While the Blackwell B200 AI GPU was initially slated for a Q4 2024 launch, SemiAnalysis’s revised roadmap suggests a low-volume shipment towards the end of the year, potentially in late November or December. The B200A, on the other hand, is expected to arrive sometime in Q2 2025 or Q3 2025.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top