Microelectronics And The AI Revolution

It is no secret that artificial intelligence and machine learning (AI/ML) are critical drivers for growth in electronics, and particularly, for semiconductors. The recent AI Hardware Summit showcased trends in AI/ML, both in enabling and using it in various application domains, including EDA. As part of the summit, Imec had organized a panel on “Advanced Microelectronics Technologies Driving the AI Revolution.” The panel included speakers from imec, TSMC, Quadric, Facebook, GrAI Matter Labs, and I joined from Cadence, representing EDA.

All of us gave brief introductions. With a balanced representation of both technology providers and users on the panel, the audience got a good summary of the challenges and needs. Required Tera-operations per second vary greatly between cloud and edge, training and inference, and drive both higher compute performance and lower power consumption. Various 3D techniques are enablers, from chip-on-wafer-on-substrate (CoWoS) through vertical stacking and wafer bonding. Imec extended the semiconductor technology challenges to the optimal mix of software, algorithms, circuits, and IP blocks, outlining how requirements change in the era of intelligent everything—mobility, health, industries, and cities. I presented how EDA enables AI/ML designs and uses AI/ML to deliver more productive flows in its incessant quest to fight design cost, which is critical in addressing design complexity. For users such as Facebook, AI/ML model complexity and power consumption were priorities. Silicon providers such as Quadric and GrAI Matter focus on deploying any AI algorithm anywhere, and eventually, enabling machines to behave naturally, saving their customers time, energy, and money.

Following the introductions, the panel addressed a couple of quite interesting questions.

AI is revolutionizing industries today by providing faster, more efficient, and accurate decision-making, from autonomous cars to virtual doctors. It is transforming the way we live, work, travel, and do business. However, is there an industry domain leading the AI revolution, while others remain fast followers? From a pure semiconductor consumption perspective, mobile, high-performance computing (HPC) seems most important. AI plays in all of them, especially, as our mobile phones become the device edge, data centers growing beyond CPU centricity with many more workload-dependent accelerators. From a pure consumption perspective, automotive and IoT may be followers.

To extend the discussion beyond pure semiconductor consumption, I highlighted Cadence’s recent consumer report, “Hyperconnectivity and You.” We asked consumers about the impact on their lives of this constant “scaling up” to make data available everywhere. More than 3,000 consumers worldwide identified mobile phones/communication as having the “highest perceived” impact, followed by health and healthcare, shopping, industrial production, military activities, and education. The common themes here are convenience, confidence, and cooperation.

Convenience is probably the primary reason consumers continue to buy connected devices and share their data so that AI/ML gives insights and suggests actions. Convenience-related features such as predictive maintenance and automated software updates will drive uptake of new hyperconnected technology, which often includes AI/ML. Removal of tedious tasks, frictionless user experiences, and clear benefits to consumers, will drive the adoption of smart-converged technology.

We also found that while consumers have confidence in technologies, they often lack confidence in data security and are reluctant to share data. This poses a challenge for all businesses in a hyperconnected data economy that applies AI/ML techniques. As the enabling industry, we must prioritize data security and deliver complete transparency on personal data usage and what the user gets back. Without user confidence, we cannot realize the full potential of hyperscale computing and the AI/ML it enables.

Finally, technology and humans need to collaborate. Hyperconnected technology employing and enabling AI/ML can make decisions and perform actions for people. However, relinquishing complete control can be scary for people. The sweet spot is people and technology working collaboratively — for example, human-led robotic surgery, or autonomous features that augment rather than replace the driver. People can feel confident in such collaborations and embrace the benefits of hyperconnectivity and AI/ML.

After a brief discussion, we agreed that automotive is probably one application domain leading the adoption of AI/ML. After all, it impacts day-to-day life significantly and it is an area where technology and humans collaborate, increasing confidence, while also delivering a more convenient experience.

The panel also discussed the most significant development challenges, including the importance of 3D integration. We emphasized that designing at and beyond the reticle limit and the need for design variations drive 3D integration and advanced node implementation, emulation, and prototyping, which address implementation and verification complexity.

Finally, we covered the technical aspects of AI/ML at the edge. As I have discussed in previous blog posts, there are many different types of “edges” to consider, and requirements on memory and performance can vary widely.

Overall, the AI HW Summit was a great experience, emphasizing several vital trends. There were also some critical new announcements – like the new On-Device Tensilica AI Platform. The panel “Advanced Microelectronic Technologies Driving the AI Revolution” organized by imec is worth a re-watch, too.