Back to the future: How analog computing can drive AI

The world is in the middle of a climate crisis, yet we are forging ahead with artificial intelligence(AI), which consumes vast amounts of energy. A paper by Alex de Vries of the VU Amsterdam School of Business and Economics, "The growing energy footprint of artificial intelligence," found that by 2027 AI systems will consume the same amount of energy as the Netherlands.

This situation is untenable. In a speech introducing the Artificial Intelligence Environmental Impacts Act of 2024, representative Anne Eshoo crystallized the situation, “AI offers incredible possibilities for our country, but that comes with high environmental costs…”

Although the proportion of electricity generated from renewables is set to rise considerably over the coming decades, focus on usage should be heat pumps and electric vehicles to bring down reliance on fossil fuels. As the world struggles with energy consumption, AI must be energy efficient, no matter how novel.

Fortunately, humans are also adept at finding solutions to problems; in this case, analog computing looks set to be the fix to the energy consumption problem of AI.

Where is the AI energy going?

It is worth looking at the figures for AI energy consumption to put the energy needs of AI into perspective. A 2023 report from the International Energy Agency (IEA), "Electricity 2024 - Analysis and forecast to 2026," highlights that electricity consumption from data centers, AI and the cryptocurrency sector will be in the range of between 620-1 050 TWh. (1 TW = a million million watts) by 2026.; by 2026, this demand is expected to double over 2022 figures.

Understanding where AI tasks consume energy helps to elucidate the problem. Generative (GenAI) is a case in point. GenAI uses Natural Language Processing (NLP), powered by deep learning that utilizes neural networks. GenAI encompasses an initial training phase that leads into an inference phase.

Neural networks are based on an architecture influenced by models of biological neurons in the brains. However, the important point to note is that an artificial neural network uses matrix multiplication math operations, which are computationally expensive.

An example demonstrates just how complex these math operations can be. The ImageNet competition of 2012 set out to explore the accuracy of neural network architectures. The "top 5 error rate" concept was used as a benchmark for deep learning accuracy. Alexnet won the most accurate algorithm competition with a top 5 error rate of 16.4%. The accuracy of Alexnet was due to the depth of the network; in other words, the number of input and output layers, each with thousands of neurons. The accuracy of Alexnet meant that the processing of a single image required 700 million different MAC multiplication and accumulation) operations, with concomitant high energy consumption.

Deep learning is energy inefficient. AI tasks that use computer vision, speech processing, and NLP require big data, and typically, the training is executed using GPUs — the processes are all energetically costly.

Energy costs of GenAI have been explored and quantified in the research paper by DeVries. The findings demonstrate the energy consumption requirements of AI training. The Big Science Large Open-Science Open-Access Multilingual (BLOOM) model consumed 433 MWh of electricity during training. Other LLMs, including GPT-3, Gopher, and Open Pre-trained Transformer (OPT), reportedly used 1,287, 1,066, and 324 MWh, respectively, for training.

The paper notes that “Alphabet’s chairman indicated in February 2023 that interacting with an LLM could ‘‘likely cost 10 times more than a standard keyword search.” The standard Google search uses 0.3 Wh of electricity.

Data from the paper shows the varying energy needed by AI-powered systems and quantifies these data more accurately:

Data source: de Vries, “The growing energy footprint of artificial intelligence”

The cost of water

An important side note is that it's not just energy efficiency that's a problem with AI; water consumption for cooling is also an issue; the heat generated by AI servers and data systems is waste heat, cooled using a supply of fresh water, that is driven by electric pumps and monitoring systems.

Google used 5.6 billion gallons of water in 2022 (mainly potable, i.e., drinking water). A 2023 Google report identified a 22% increase in water consumption, driven mainly by AI.

Clearly, this resource consumption is unsustainable as AI models evolve in complexity. However, industry experts are looking back to analog computing to find a fix for energy-hungry AI.

Overview of analog vs digital computing

Real world information is analog in form - sounds, temperature, height, etc.

Analog computers can perform computational operations on direct representations of these continuously variable data. For example, using a variable voltage to represent temperatures in a temperature controller, or even using water flow to represent economic variables (as in MONIAC (Monetary National Income Analogue Computer, used to predict economic futures).

Conversely, digital computers represent and process data as discrete values, represented as strings binary digits (bits) of zero and one.

Digital computing has overtaken analog computers because this method of processing information offers significant advantages. Although theoretically analog computers have infinite resolution, in practice resolution is limited by a combination of the maximum and minimum values that can be processed directly (usually a voltage, say 0 V to 10 V) and resulting inherent electrical noise, which means that precision is limited in analog, perhaps to only 1 in 10,000. The binary nature of digital computers means they are effectively immune to electrical noise, and the dynamic range of numbers is determined by the number of bits, effectively limitless, thus affording extremely high precision. Programmability is an inherent quality of digital computers, and one that is not usually straightforward in analog computing. In classical analog computers the program is set by patching connections and setting dials. Digital computers are programmable by design, using easily written and editable software programs. Next up is memory; it is more difficult to reliably store analog values in a format that is quickly retrievable. In contrast, high-speed digital memory is readily available.

However, analog computing still has advantages over digital, including speed, as analog computing is essentially instantaneous, whereas digital relies on a clock to step through instructions; therefore, processing speed is inherently limited by the clock rate. In addition, when processing real world data in real time, the speed of analog to digital conversion is also significant. Importantly, the energy required in analog computing is significantly smaller than digital: digital computing typically uses thousands of transistors to perform even simple calculations and the switching of these transistors is the main source of power consumption. (Consequently, higher clock speeds increase power consumption.).

Analog computers may seem to have been relegated to the past, but in recent years, they have found new fandom in areas as diverse as simulating biological organisms and applications in control engineering. Now, analog looks to be an enabler for AI by optimizing energy consumption.

Is analog the answer?

Analog computing has not disappeared, and its intrinsic value of speed and energy efficiency is the perfect match for the programmability and precision of digital computing. This perfect pairing is now behind many research projects and commercial ventures exploring how to optimize energy use in AI programs.

Analog computing can be used in the training and inference phases to reduce energy consumption. In training, the "von Neumann bottleneck" is a phenomenon that describes how delays are caused by using the same channel to fetch data and instructions; the CPU waits for memory to be released to process instructions. This duality reduces operating efficiency. Companies like IBM are exploring analog solutions to this bottleneck.

Resistance memory is a technology required to make analog-based AI even more compelling. It means that during the training phase, you can adjust the input weighting directly in analog format as a resistance rather than retrieving/storing a digital value in memory. A specific voltage or current may represent analog processing of data values. To set a value (i.e., set the voltage or current), users set a resistor value according to Ohm's law [V = IR]. However, for this to be programmable, operators need a variable resistor - in this case, one that varies its resistance with an applied control voltage.

Inference is also computational (and energy) inefficient. Analog AI chips use far less energy. Startups are creating analog chips that run neural networks. Mythic AI is one such company. Mythic AI provides analog chips that can perform 25 trillion math operations per second using 3 W. The digital equivalent would use up to 100 W (and is much more expensive).

Another example of energy efficiency is using analog systems to listen for the wake word in Smart AI systems, like Alexa. The analog system is then used to switch on the circuitry of the device, lowering power needs.

Work is ongoing to reduce the computational parameters to optimize energy efficiency significantly. Much of this research focuses on improving network structure efficiency. An example is the Energy Efficient Deep Neural Network (EDEN) architecture. EDEN uses a process called "approximate memory." This reduces reliability, which leads to high bit errors; however, EDEN also reduces energy consumption. EDEN leverages neural networks' ability to tolerate bit errors, thereby maintaining reliability and decreasing energy needs.

A software programmable processor offers a hybrid state. Analog and digital are used symbiotically to optimize the processing and the required energy. For example, in a neural network with multiple layers, you'd use digital precision to avoid the noise inherent in analog. In this case, the analog output is converted to digital before being sent to the next block in the neural net, which is then converted back to analog to preserve the signal, and so on. This symbiotic hybrid may hold the key to optimizing AI systems in a world where energy is at a premium.

The future of AI is hybrid

Sometimes, the past provides answers to present and future problems. After all, ideas and concepts don't always follow linear patterns. We may think of digital technology as the modern way to process information, and this may be true, but what if the price of energy in a world in crisis is just too high?

Referring to the paper by De Vries, he notes that the applications of AI will be a determining factor in energy consumption, concluding: “…it would be advisable for developers not only to focus on optimizing AI, but also to critically consider the necessity of using AI in the first place.”

This refocusing on the place of technology in society is critical. Technology, for technology's sake, could cost the earth. However, important socio-technical advances can be saved from extinction by optimizing how technology is used, even if this means delving into computing's past. Digital computing is at a juncture; the miniaturization offered by turning to digital is reaching its limits (Moore's Law). Analog looks set to ensure that AI does not hit an energy wall, too.

About the author

Susan Morrow, originally a scientist, has been working in the IT security sector since the 1990s. She has developed expertise in various areas such as encryption, digital rights management, privacy, online identity and emerging technologies. Susan has contributed to government, enterprise, consumer security and identity projects. She was recognized as one of the most influential women in U.K. technology by Computer Weekly in 2020, 2021, 2022 and 2023, and was also shortlisted in the top "100 Women in Tech" in 2021.