Guo Mingchi: NVIDIA's next-generation AI chip R100 is expected to start mass production in the fourth quarter of next year, using TSMC's N3 process

Author: Zhao Ying

Source: Hard AI

On May 8th, Tianfeng International Securities analyst Guo Mingchi released an updated forecast, indicating that Nvidia's next-generation AI chip R series/R100 AI chip may start mass production in the fourth quarter of 2025, with the system/rack solutions expected to start mass production in the first half of 2026.

Guo Mingchi pointed out that the R100 will adopt TSMC's N3 process and CoWoS-L packaging, and is expected to be paired with 8 HBM4.

In comparison, the "most powerful AI chip" released by Nvidia in March, the B100, adopts the N4P process technology and also uses CoWoS-L packaging. The B100 marks a 1000-fold increase in Nvidia's AI chip computing power in just 8 years.

The Grace CPU in the GR200 integrated system will use TSMC's N3 process, while the Grace CPUs in GH200 and GB200 will use TSMC's N5 process.

Furthermore, the R100 will adopt approximately 4 times the reticle design (compared to 3.3x for B100), and the size of R100's interposer has not been finalized, with 2-3 options available.

It is worth mentioning that energy efficiency is also a key focus of the design. Guo Mingchi stated:

"Nvidia has realized that power consumption of AI servers has become a challenge for customer procurement and data center construction, so the chip and system solutions of the R series not only focus on improving AI computing power, but also on improving power consumption."

In the previous generation of chip design, compared to H100, the cost and energy consumption of GB200 have been reduced by 25 times. Huang Renxun stated that previously, training a 18 trillion parameter model required 8000 Hopper GPUs and consumed 15 MW of power. But now, this goal can be achieved with 2000 Blackwell GPUs, consuming only 4MW of power