NVIDIA’s highly anticipated Blackwell AI GPUs have encountered major setbacks due to design flaws, impacting the production and launch of its new B100 and B200 models. According to a report by SemiAnalysis, these issues stem from the Blackwell die itself, potentially requiring a redesign. Additionally, the bridge dies within TSMC’s advanced CoWoS-L packaging technology also present challenges.
These complications have forced NVIDIA to adjust its roadmap. The B200, initially set for a Q4 2024 release, will now see limited shipments in the latter part of the year. The company is now focusing on the B200A, a revised version of the GPU, which is expected to arrive in the second half of 2025. This new model utilizes the B102 die, also used in the China-specific B20 Blackwell GPU. The B102 features a single monolithic compute die with four stacks of HBM and can be packaged using TSMC’s CoWoS-S technology, or even other 2.5D packaging suppliers like Amkor, ASE, and Samsung.
The B200A will be available in 700W and 1000W HGX form factors, offering up to 144GB of HBM3E memory and a memory bandwidth of up to 4TB/sec, though this is slightly lower than the H200. To mitigate the impact of these delays, NVIDIA is extending the Hopper AI GPU series and introducing the B200A as a new addition to the Blackwell family. This strategic move aims to address the challenges encountered with the initial Blackwell design while ensuring a continued flow of advanced AI computing solutions for the market.
