Semiconductor Chip Design in a Legoland

Moore's Law and Dennard scaling have been ticking and propelling the semiconductor industry for decades. They have also pushed the digitization and pixelization of our everyday life. The industry has attained a great achievement in terms of Power consumption, Performance, silicon Area, chip and system Cost as well as Time to market (PPACT). Unfortunately, the pace and benefits of CMOS scaling have now slowed down, yet the explosive demand for chip PPACT is on a completely opposite trajectory due to the rise of Generative AI (GAI), next-generation wireline and wireless (5G-Advanced and 6G) communication, automotive electronics and so forth. On the energy supply side, the worldwide electricity generation increasing at ~6%/year is far less than the growth of the power consumption for just computing alone [1] making the supply vs demand unsustainable. Emerging technologies, such as new material, compound semiconductor devices, and novel circuits and systems both in the Cloud and at the edge are coming to the rescue. But splitting functions then stitching diverse chiplets together is still at the dawn of the new homogeneous and heterogeneous integration (HI) era. It should also be noted that HI should not be limited to semiconductor chips in a package. Some promising progress will be illustrated ranging from electrical, optical, and passive component integration to support co-packaged optics (CPO), Antenna-in-Package/Module (AiP/AiM), socket waveguides, etc. While the advance in devices and manufacturing is essential, there are lots of innovations needed for architectures, systems, algorithms, and circuits to develop the huge variety of applications. Tools like Computer-Aided Design (CAD), Electronic Design Automation (EDA) and automatic software code generation are supposed to help human engineers to find optimal solutions in a timely manner. However, the pace of enhancement for those tools is around an order of magnitude slower than the Moore's Law. The emergence of machine learning and inference, commonly called ML or AI, is showing a great potential [2], [3] especially when the engineers need to deal with high-dimensional complex options. In fact, many designs assisted by ML have surpassed the PPACT achieved by human engineers using conventional software tools. Following the new large AI foundation model framework, “ChipGPT” could become an essential tool in the future. So far, we have only cracked the tip of the iceberg for all the aforementioned fronts. A lot more challenges and opportunities still lie ahead to be explored.