NVIDIA's new Blackwell AI GPUs 'encountering major issues' which require redesign, big delays

NVIDIA's upcoming Blackwell AI GPU family is "encountering major issues" according to a new detailed breakdown by SemiAnalysis, which will impact production targets through into the first half of 2025.

3

The main problems are from the NVIDIA Blackwell GPU architecture and the chip packaging used: TSMC's new CoWoS-L. SemiAnalysis explains that one of the rumors swirling is that one of the top few global routing metal layers and bump out of the Blackwell die needs a redesign, while another is that the bridge dies inside the CoWoS-L packaging need a redesign.

SemiAnalysis explains that there has a been a warpage in the CoWoS-L packaging that was caused by heat, with TSMC's new CoWoS-L advanced packaging being more complex than the now more mainstream CoWoS-S packaging.

NVIDIA's new Blackwell AI GPUs will be the first mass-market chips to use TSMC's latest CoWoS-L advanced packaging, with "NVIDIA and TSMC aimed for a very aggressive ramp scheduled to over a million chips a quarter. Consequently, there have been a variety of issues".

• Read more: NVIDIA's next-gen Blackwell AI GPUs delayed, rumor has it 'design flaws' are to blame

According to SemiAnalysis, NVIDIA is reportedly dealing with the issues by extending the Hopper AI GPU series and introducing a new Blackwell family GPU: the B200A. The new B200A AI GPU would be packaged using TSMC's CoWoS-S and rival packaging technologies from other suppliers, including Amkor, ASE, and Samsung.

According to SemiAnalysis, the new B200 AI GPU will fulfill demand for lower-end and mid-range AI systems, replacing the B100 and B200 chips in the HGX 8-GPU supercomputing platform form factors.

• Read more: TSMC to expand CoWoS equipment orders: huge AI chip demand from NVIDIA, Amazon, Meta, Tesla
• Read more: NVIDIA's new Blackwell AI GPUs drives TSMC to increase CoWoS output by over 150% in 2024
• Read more: TSMC is already buying equipment for two CoWoS advanced packaging plants to be built in Taiwan

TSMC has been aggressively expanding its CoWoS-L advanced packaging capacity to boost volumes, with a new packaging plant called "AP6" now under construction, and by converting current CoWoS-S capacity at AP3 to CoWoS-L capacity, added SemiAnalysis.

3

NVIDIA's new Blackwell B200 AI GPU was meant to be launching in Q4 2024, but SemiAnalysis has a new roadmap that has B200 shipping in "low volume" towards the end of the year (late November, December) while B200A will arrive sometime in Q2 2025 or Q3 2025.