I was doing some history reading on computing earlier this week and got hooked on the concept of analog computing. It’s fascinating to think about how this technology, that dates back several decades, is making a comeback even in the AI era.
First, the basics. Digital computing is all about 1s and 0s. Meaning you have an electrical signal and you convert it into digital domain via an ADC(Analog to Digital Converter) and do the rest of the process in mostly software. Although it has many good things going for it like its precise, configurable & repeatable, one of the main points is its slow(think conversion time) and power-hungry because of the extra steps needed.
This is where analog computing shines. Unlike digital, analog doesn’t use binary. It uses continuous signals, like voltages or currents, to represent and process information. It’s fluid, parallel, and highly efficient. For AI tasks, like neural network operations that involve lots of matrix math, analog systems can process data directly without all the energy-intensive conversions digital systems need. Its energy efficient and can perform computations at a fraction of the cost making them great for AI edge applications like sensors, cameras, wearables etc.
A company worth looking into is Mythic AI, which uses compute-in-memory technology. Here, matrix multiplications happen directly in the circuit, using analog signals. Imagine a DAC generating voltages across varying resistors; measuring the current in the line gives the multiplied result V=IR. This is a fundamental multiplication block. Scale this across a large node matrix, and you achieve fast, low-energy matrix operations without transferring data between memory and processor which cant be avoided in digital computing.
I think the future might be about combining the two to create systems that are both powerful and efficient. As of today, since the o3 release from OpenAI, I feel the only wall(if any) AI is going to hit, is the compute shortage wall, nothing else.
