Nvidia's next-generation R100 GPU uses TSMC 3nm process with CoWoS-L packaging and next-gen HBM4 shipping in Q4 2025.

NVIDIA's next-generation R-series and R100 AI GPU will enter mass production in Q4 2025, with the system and rack solution following in Q1 2026. According to industry insider Ming-Chi Kuo, the R100 will be manufactured on TSMC's N3 process node and use CoWoS-L packaging.

The R100 AI GPU features a 4x reticle design, compared to the B100's 3.3x reticle design. The chip will include 8 x HBM4 units, while the interposer size for the R100 has yet to be finalized, with Ming-Chi noting there are 2-3 options under consideration. NVIDIA's accompanying GR200 Grace CPU will also use the TSMC N3 process, a step up from the N5 process used in the GH200 and GB200 Grace CPUs.

While the B100 currently uses TSMC's N4P process node, the R100 represents a transition to the newer N3 technology. Both generations will utilize CoWoS-L packaging, maintaining consistency in the packaging approach across these generations.

Ming-Chi notes that NVIDIA has recognized power consumption in AI servers as a significant challenge for customers' procurement and data center construction decisions. As a result, NVIDIA's next-generation R-series AI chips and systems are designed with a dual focus: advancing power efficiency alongside delivering the next levels of AI computing capability.