Advancing Microelectronics to Meet AI's Evolving Demands

Artificial intelligence is redefining technology across industries, demanding microelectronic solutions that deliver unprecedented performance, energy efficiency, and compact form factors. From massive data centers to autonomous vehicles, robots, mobile devices, wearables, and future applications yet to be imagined, the need for faster, smarter, and greener hardware is pressing.

In the hardware realm, this challenge calls for innovative paradigms in sensors, processors, memory, interconnects, and packaging. Leti’s research community is actively exploring these frontiers, positioning itself at the intersection of Edge AI and broader industry trends. Real‑world, manufacturable solutions will emerge only through interdisciplinary collaboration before mass production.

We can anticipate the AI marketplace by mapping applications across computing capability and power consumption (see Figure 1). Wearables sit at the low‑power, low‑compute end, while data centers demand the highest performance and energy budgets. Smart appliances, augmented reality, robotics, and autonomous vehicles occupy the spectrum in between.

Figure 1. (Source: Leti)

Edge AI—where most data analysis occurs at the point of capture—thrives on this left‑hand side of the spectrum. Achieving it requires sensors and processors that can adapt on the fly, much like human eyes and ears. These components must dynamically adjust characteristics such as dynamic range to match local intelligence, all within a footprint that fits a watch or a wearable.

Conversely, larger‑scale AI workloads expose the limits of traditional computing architectures, especially the constant memory read/write cycles that drain both time and energy.

With these realities in mind, Leti prioritizes research into smart sensors and groundbreaking computing approaches.

One of modern computing’s core bottlenecks is data movement: transferring data between memory and processor now consumes far more time and energy than the computation itself. In fact, data transfer and memory access can account for up to 90% of a system’s energy budget—an alarming figure for AI workloads that rely on vast datasets and simple operations. Reducing data movement is therefore essential.

Leti’s long‑standing work on 3D circuitry seeks to stack memory directly onto processors, shortening physical links and cutting latency. We are also pioneering new memory architectures that perform addition, subtraction, and Boolean logic directly within SRAM cells. This in‑memory computing (IMC) approach eliminates the need to shuttle data out of memory, saving space and energy. Early prototypes show the potential for a 100‑fold increase in AI throughput while staying within the same frequency and energy envelope expected for the 2020s.