TSMC: CoW-SoW technology will change the future of large chips by 2027 | GPU hardware | Gpu comparison | Nvidia hardware | Turtles AI

TSMC, a global leader in semiconductor manufacturing, is working on an advanced packaging technology called CoW-SoW (Chip-on-Wafer, System-on-Wafer) that is expected to enter mass production by 2027. This innovation aims to meet the growing needs of the AI and data center sectors, where the demand for high performance and large-scale integration capabilities is constantly increasing.

Key points:

• CoW-SoW technology: TSMC develops a new packaging technique for large-scale chips, expected in 2027.
• Focus on AI and Data Center: The technology is targeted at optimizing chips for AI and data center infrastructure.
• Impact on manufacturers: Companies such as NVIDIA are forced to revise their architectures due to manufacturing complexities.
• HBM capacity: CoW-SoW promises a significant increase in memory capacity and overall chip performance.

SMC is planning to revolutionize the semiconductor industry with the introduction of its new “Chip-on-Wafer System-on-Wafer” (CoW-SoW) packaging technology, expected for large-scale production by 2027. This technology is designed to address the growing needs for computational capacity, particularly in the AI and data center sectors. CoW-SoW will allow logic and memory chips to be stacked on a single interface, overcoming current limitations in bandwidth and performance, and enabling the development of chips of unprecedented size.

TSMC has already shown a preview of what will be possible with this technology, stating that CoW-SoW will offer up to 60 times the integration capacity of current techniques. This will be especially useful for future data centers, where efficiency and speed of computation are essential. The project also aims to solve some of the problems encountered with NVIDIA’s Blackwell architecture, which has been delayed due to complexity in manufacturing large chips.

This new technology is an evolution of the previous InFO-SoW, which already allowed the integration of wafer-scale processors but was limited by the use of on-chip memory. With CoW-SoW, it will be possible to integrate HBM4 memory directly onto wafers, further improving the performance and processing capability of AI and HPC systems.

However, manufacturing these chips will also bring new challenges, especially in terms of heat management and power consumption. Large chips will require advanced cooling solutions, such as liquid or immersion cooling, to maintain performance without compromising the physical integrity of the components.

 CoW-SoW enables the integration of memory and logic chips on a single interface, improving interconnect speed and accuracy when connecting different dies. TSMC’s proposed solution is not only a response to current needs, but also a step forward in addressing the technical challenges that arise with the development of increasingly complex chips, such as those used in NVIDIA’s new architectures.

NVIDIA has recently faced significant challenges with its Blackwell architecture, which is currently the largest AI chip on the market. Manufacturing complexities, especially related to the interconnection between the GPU die and the substrate, have forced the company to revise its designs, causing production delays. TSMC’s own CoW-SoW technology, already in use with Cerebras’ wafer-scale chip, has proven to be a promising solution to overcome these difficulties, enabling higher HBM memory capacity and higher lattice density than current technologies. In particular, this technology focuses on optimization for data center clusters, making it particularly suitable for large-scale AI applications.

As TSMC continues to refine CoW-SoW, NVIDIA and other chipmakers are looking with interest at this solution, which could be integrated into future architectures, such as NVIDIA’s next-generation Rubin. Ultimately, the ability to produce chips of increasing size and complexity with superior efficiency could represent a significant shift for the semiconductor industry, pushing the boundaries of what is possible in the fields of artificial intelligence and data centers.

TSMC’s innovation, with its prior experience in large-scale chip management, thus looks set to play a key role in the future of semiconductor technology.