The rapid rise of GenAI development has significantly accelerated the pace of innovation within the industry. There are reports that NVIDIA has contacted TSMC to provide it with the next-generation "Blackwell" B100AIGPU based on TSMC's 3nm process. The next generation Blackwell GB100 GPU uses a chiplet design with significant changes.
Visit the purchase page:
JD.com NVIDIA series product summary
Before we delve into the report, we need to review the Blackwell GB100 GPU. They will be TeamGreen's first AIGPUs designed with small chips and are expected to bring decisive improvements in performance per watt. There have been reports in the past that the GB100 will feature a more complex "structural design," which could lead to lower production volumes at the outset.
It was rumored that the Blackwell GB100 GPU will use TSMC's 3nm process, and now it seems that the deal has been confirmed. However, currently TSMC, especially its 3nm manufacturing, faces obstacles in terms of utilization and lack of advanced manufacturing plants. Therefore, current actual implementation of 3nm may not be possible until 2024, when the Blackwell GB100 is scheduled to be released.
The DigiTimes report once again makes it clear that NVIDIA is making a full push into the artificial intelligence industry, even preventing competitors such as Intel and AMD from gaining a foothold. , and the next-generation Blackwell GB100 AIGPU may play a crucial role here, as it will mark the transition of the AI industry to next-generation development.
AMD's EPYC Instinct MI300 accelerator is expected to be launched on the 5nm process node later this year, so the follow-up is expected to be launched around 2025-2026, which means Blackwell GPU will lead on the advanced process node. Intel is expected to launch the Gaudi3 artificial intelligence accelerator based on the 5nm process node next year, and launch FalconShores follow-up products in 2025.
NVIDIA is currently in the driver's seat, with the company aiming to generate up to $300 billion in revenue from artificial intelligence by 2027, so it can't afford to let up any slack.