Nvidia's Rubin architecture GPU adopts small-chip technology, and TSMC has been of great help
At the previous GTC2025 conference, NVIDIA CEO Jensen Huang unveiled the latest technical roadmap blueprint, which has aroused great anticipation in the industry. The latest news is that TSMC will collaborate with NVIDIA to develop the next-generation advanced small-chip GPU, which will work in conjunction with NVIDIA's "Rubin" architecture and serve as the successor to the Blackwell architecture.
According to the foreign media Digital Trends, the difference between small chips and traditional single-chip Gpus is obvious, with higher performance, scalability and cost efficiency. Small chips enable chipmakers to package multiple smaller semiconductor chips into a single chip, increasing output and reducing production costs.
This is becoming increasingly popular in the semiconductor industry. Chip design is becoming more and more complex, and the limitations of traditional process scaling are getting larger and larger. With TSMC's packaging process, NVIDIA can improve the energy efficiency and performance of Gpus, which is very suitable for artificial intelligence, data centers, and high-performance computing, etc.
Nvidia Rubin architecture GPU adopts TSMC's N3P process and is the optimized version of TSMC's 3-nanometer family. Compared with the previous generation, it has improved performance, power efficiency and transistor density, maximizing the advantages of the chip. This enables NVIDIA to maintain energy efficiency while breaking through the performance limits of the GPU.
To optimize GPU performance, NVIDIA also adopts TSMC's advanced packaging such as SoIC, with chips stacked vertically to enhance power efficiency and reduce transmission latency between GPU chips.
TSMC intends to increase its advanced packaging capacity and plans to expand its SoIC capacity by the end of the year. The NVIDIA Rubin architecture series adopts SoIC, which can fully utilize the functions of HBM4. The Vera Rubin NVL144 platform has two standard-sized Rubin Gpus with FP4 performance of up to 50 pflops and 288GB of next-generation HBM4. The higher-end NVL576 adopts Rubin Ultra Gpus with four standard-sized chips, with the performance enhanced to 100 pflops of FP4, accommodating 1TB HBM4e on 16 HBM stacks.
Nvidia's adoption of small chips is in line with the industry trend, while AMD and Intel, among others, have integrated similar designs into their processors. The modular feature of the chip enables chip manufacturers to mix different processing units and optimize specific working performance. Artificial intelligence and high-performance computing are driving a strong demand for hardware. The collaboration between TSMC and NVIDIA on GPU is expected to achieve breakthrough progress.
BEIJING 
ENGLISH 
USA 
DEUTSCH 
FRENCH 
JAPAN 
SPANISH 
AUSTRALIA
