Layout Generators For Artificial Intelligence Hardware Design

Artificial intelligence (AI) is a powerful tool that offers great convenience in many areas of life. In addition to improving Internet searches and online shopping, it enables driver assistance systems that can save lives, for example. AI in its various forms is the essential tool for such applications, and it can be expected to show a similar development as microelectronics did.

Although AI often appears to be embedded in the end user device, such as a smartphone, the required computing power is generally only available on servers in the cloud. This means the Internet must always be available, and the latencies involved can be relatively long. In particular, signal analysis applications with extremely low energy consumption and extremely long battery lifetimes or time-critical applications cannot rely on such centralized architectures. In these cases, the complex signal analysis must be performed locally at the edge, in distributed fashion, making special hardware indispensable.

In the past, typical hardware solutions for such signal analysis problems utilized ASICs (application-specific integrated circuits) with classical digital signal processing. Particularly for complex and highly complicated requirements that are difficult to program, these conventional hardware architectures can no longer keep up. New AI methods are required in order to achieve the required energy efficiency and/or analysis quality on the edge devices. With the transition from typical digital signal processing to the new AI methods, AI is also being implemented in digital circuits. However, frequent matrix operations can quickly become a bottleneck here. Analog crossbar arrays have the potential to solve this problem while improving energy efficiency and latency.

Yet the analog design process poses a challenge in terms of design time, costs and associated risk, especially since the automatic layout synthesis available for digital circuits is not available for analog circuits. The designers must manually design much of the highly complex and complicated circuits. When it comes to the regular arrangements typical of crossbar designs, this can become a painful task. Modern layout generators are an appropriate solution for overcoming these analog design challenges. Generators are parameterizable programs that describe the target layout, allowing it to be generated automatically. They employ routines – controlled by the designer – to produce layouts in just seconds or minutes, which the designer can use directly in the design like Lego bricks. Complex and, in particular, regular layouts can be quickly, effectively and reliably generated in this way. Especially for extremely customized layouts, this process is based on close cooperation between designers and generator developers in order to create the right generator for the design problem. This cooperation makes it possible to efficiently develop highly customized circuits by means of automation.

These generators then perform similarly to memory compilers, which also generate complex, customized layouts based on specified parameters. In the context of AI hardware, however, entirely new possibilities arise: The development process begins with an application analysis to partition the system into classical signal processing and AI methods, followed by algorithm design and related parameterization of the appropriate generator for automated layout design. This systematic approach coupled with the use of layout generators opens up the possibility of developing high-performance signal-processing applications based on AI hardware more quickly, affordably and reliably.

With generators as AI hardware design tools, it is possible to economically meet the challenges posed by the complex and variable requirements of modern applications. Just as design automation became widespread for digital design, analog design automation is now ready to overcome critical bottlenecks, allowing projects to be completed more quickly and cost-efficiently.

When will you generate your hardware? Find out more information on analog generators.