Blog

May 17, 2023

Analog microchips are the cure for AIs lust for power

visualspace/iStock By subscribing, you agree to our Terms of Use and Policies

visualspace/iStock

By subscribing, you agree to our Terms of Use and Policies You may unsubscribe at any time.

In the past, analog chips governed computing, operating over continuous value ranges. While analog computing may appear obsolete compared to today's standards, lacking the precision and adaptability of digital chips, it is experiencing a revival in the realm of things like advanced AI. But what are they, and how could they be useful?

Let's find out.

An analog chip, or analog-integrated circuit (IC) or device, is a semiconductor device that processes and manipulates analog signals. Analog signals are continuous, time-varying electrical signals representing real-world phenomena like sound, temperature, pressure, and light. These chips are used in many applications, including consumer electronics, telecommunications, automotive systems, and industrial automation.

Put another way; analog chips work somewhat similarly to the cells of your brain. Unlike digital chips, which use 1s and 0s to process information, analog chips work with a continuous range of values, like a dimmer switch that can be adjusted smoothly.

Analog chips are characterized by their ability to handle continuous signals, as opposed to digital chips that work with discrete, binary signals (0s and 1s). The fundamental components of analog chips include transistors, capacitors, resistors, and diodes, which are interconnected to create specific circuit functions like amplification, filtering, and signal conversion. For this reason, analog chips enable seamless interaction between the natural world and the digital domain, ensuring that our electronic systems remain efficient, reliable, and versatile.

Despite the increasing prevalence of digital technology, analog chips remain indispensable due to their unique ability to handle real-world signals. The development of mixed-signal ICs, which combine analog and digital circuitry on a single chip, has further increased the importance of analog design and expertise. This integration allows for more compact, efficient, and cost-effective electronic systems, fueling innovation in numerous fields.

One of the most common types of analog chips is the operational amplifier (op-amp). Op-amps are versatile components that perform various functions, such as signal amplification, filtering, and mathematical operations. They are integral to many electronic systems, from audio equipment to medical devices.

Teravolt/Wikimedia Commons

Other notable examples of analog chips are the analog-to-digital converter (ADC) and the digital-to-analog converter (DAC). These chips convert continuous analog signals into discrete digital data and vice versa, enabling communication between the analog and digital domains. ADCs and DACs are critical in converting digital signal processing, data acquisition, and digital audio and video systems.

Voltage regulators are another important type of analog chip. They provide stable voltage levels to various electronic circuits, ensuring the system operates reliably and efficiently. Voltage regulators are used in multiple applications, from power supplies and battery chargers to automotive electronics and renewable energy systems.

Analog chips also feature heavily in specialized devices tailored to specific applications, such as sensor interfaces, radio frequency (RF) circuits, and power management ICs. Sensor interfaces enable the processing and conversion of signals generated by sensors like temperature, pressure, or light sensors. Radio frequency circuits facilitate wireless communication in cell phones, Wi-Fi routers, and satellite systems. Power management ICs, on the other hand, regulate and distribute power within electronic systems, optimizing energy consumption and prolonging battery life.

Analog and digital microchips differ in functionality, signal processing, and application areas—one of the most important differences is how signal processing occurs with each kind of chip. Analog microchips process continuous, time-varying electrical signals representing real-world phenomena like sound, temperature, and light. In contrast, digital microchips work with discrete, binary signals represented by 0s and 1s. These binary signals perform logical and arithmetic operations in digital systems.

Another notable difference is that analog microchips are designed to handle specific functions such as signal amplification, filtering, and conversion. Examples of analog microchips include operational amplifiers, analog-to-digital converters, and voltage regulators. Digital microchips, on the other hand, can execute complex logic and arithmetic operations based on binary data. Examples of digital microchips include microprocessors, microcontrollers, and memory chips.

Applications that require continuous signal processing or control rely on analog microchips, such as those found in audio equipment, sensor interfaces, and power management systems. For data processing, storage, and communication, digital microchips are crucial, as seen in computers, smartphones, and digital communication systems.

Circuit components make a difference as well. Analog microchips are comprised of fundamental components such as transistors, capacitors, resistors, and diodes, which are interconnected to create specific circuit functions. Digital microchips, on the other hand, consist of digital logic gates built from transistors, which are organized into more complex structures like flip-flops, registers, and arithmetic logic units.

Noise susceptibility is another distinguishing factor. Analog microchips are more susceptible to 'noise' (small, undesired variations in voltage) and signal degradation, as even small variations in signal levels can affect the system's overall performance. Digital microchips are less sensitive to noise because they operate on discrete voltage levels, making it easier to distinguish between 0s and 1s. However, analog-to-digital and digital-to-analog conversion processes may introduce noise and quantization errors in mixed-signal systems.

Andrzej Barabasz/Wikimedia Commons

Today's AI mainly works using the mathematical operation of matrix multiplication, which involves multiplying two rectangular arrays of numbers. This occurs when information travels between artificial brain cells or neurons. To make fast decisions, many 'artificial neurons' must simultaneously send data to many other neurons. This is a huge, complicated process. For this reason, graphics processing units (GPUs), made for handling these fast, big processes, are now widely used in AI development. Powerful GPUs and specialized AI chips make it possible to create bigger neural networks with tons of virtual neurons processed by thousands of GPUs. While this is great for AI research and improving AI's abilities, it has some problems.

Today's GPUs have billions of tiny transistors that use a lot of power and make a lot of heat. Think about how much power and heat thousands of GPUs would make for training one AI, as well as the power needed to keep the equipment cool. It would be like using more power than several houses in a year and would take up a lot of space.

In fact, environmentalists and others have frequently pointed out that AI already has a huge environmental impact, and this will grow even larger should it become more widespread. It has been estimated that training just one AI model has the around the same environmental impact as the lifetime output of five cars – including their manufacture.

When an AI is ready to do its job, it still needs a big GPU with all the power and heat it creates. This also makes it more difficult to put AI into small devices like cameras or robots, where there is little room for all that power and heat.

And this is where analog chips could help. While less capable in some areas than digital chips, analog chips are fast and use less power for just one task, like multiplying big groups of numbers. This makes them good competition for digital chips. A special analog AI chip uses less than 10 watts of power, while a GPU uses more than 100 watts for the same thing. This means we can put AI into smaller devices where power and heat matter, like an AI-powered camera on a factory line, which can recognize parts without sending tons of data to another system and waiting for an answer.

Analog chips can't replace digital ones for everything AI does, especially when working with people or getting info. But combining the best parts of both types of chips could help make AI even better, not just for high-tech stuff but also for smart devices in homes and factories worldwide.

That being said, some companies are making huge strides in this area. Hardware startups like Innatera, Rain Neuromorphic, and others are creating analog chips with neuromorphic circuitry to mimic brain functionality for AI computing. The brain is inherently analog in how it takes in raw sensory data, and these chipmakers aim to recreate how neurons and synapses work using traditional analog circuitry. Analyst Kevin Krewell of Tirias Research noted that analog chips are particularly suitable for low-power sensing devices, especially for sound and vision applications.

AI and machine learning primarily rely on digital chips, but there is a place for analog chips on the edge, such as in smartphones and cars, which require instant intelligence without needing to send data to the cloud. Innatera's third-generation AI chip has 256 neurons and 65,000 synapses and runs inference at under one milliwatt. The chip is used by customers to run radar and audio applications, aiming to incorporate low levels of learning and inference on the device.

Innatera's chip takes information from a sensor, converts it into spikes, and encodes the content based on when these spikes occur. The chip mimics the brain by scaling the current going into and coming out of the artificial neuron. The goal is not to disrupt the AI flow into the cloud but to replace AI chips on the edge that cannot currently make on-device decisions. The chip reduces the process of converting analog signals to digital by converting analog signals into spikes.

Analog circuits have limitations, such as difficulty scaling like digital circuits and requiring conversion to digital to interact with the rest of the system. Despite these challenges, neuromorphic chips have the potential to provide on-device intelligence with better power efficiency compared to AI in the cloud.

And that is your lot for today.

While digital chips are, by far, the most dominant form on the market, their high energy consumption (and heat production) are a significant limiting factor for them if computing-power-thirsty processes like AI is to become widespread in society. By hard-coding (i.e., printing), simple or more complex AI algorithms onto analog chips might be the cure.