In the future, computing devices are poised to become almost frighteningly capable. They will be able to calculate trillions of digits of pi and substitute for humans in numerous capacities. Yet, even now, humans possess an internal computer far surpassing any existing AI, consuming merely a fraction of a percent of the energy.
“The human brain is an amazingly energy-efficient piece of hardware,” remarks Advait Madhavan from the National Institute of Standards and Technology (NIST). “In computational terms, it can achieve the equivalent of an exaflop—a billion billion (1 followed by 18 zeros) mathematical operations per second—using just 20 watts of power.”
Viewing the brain as an organ, it is undeniably energy-intensive: constituting roughly two percent of body mass, it accounts for about 20 percent of our basal energy expenditure. However, objectively, this isn’t excessive; with an average daily intake of, say, 2700 calories, the brain requires only about 340 calories to fuel itself.
This translates to 0.4 kilowatt-hours—enough energy to keep an old 60-watt incandescent bulb lit for less than seven hours. That’s less energy than one might derive from, for example, three bananas. Fundamentally, it’s a negligible amount, especially when contrasted with the alternative: “One of the world’s foremost supercomputers, Oak Ridge Frontier, recently demonstrated exaflop-level computation,” Madhavan noted. “But it demands a million times more energy than the brain—20 megawatts—to perform that feat.”
Consequently, running the Oak Ridge Frontier power plant would necessitate energy equivalent to, perhaps, three million bananas. A more sensible comparison, though, is this: sustaining the facility for one day would require burning 207 metric tons of coal, emitting 340 metric tons of CO₂. Alternatively, 120,000 liters of crude oil would yield roughly half that CO₂ output; yet another option is burning 84 million liters of natural gas, resulting in only 74 metric tons of CO₂ emissions.
These expenditures are not indefinitely sustainable. However, the trajectory of computers and artificial intelligence—with their ever-increasing energy demands—shows no signs of reversing soon. At this juncture, it becomes pertinent to ask: Can we devise better methods?
There is a frequently cited—though untrue, yet persistent—tidbit that humans use only 10 percent of their brains at any given time. While in reality, we employ most of our brain most of the time, and nearly all of it some of the time, this assertion hints at a more subtle truth.
“When considering a healthy human brain, it does not engage all its neurons nor utilize its full cognitive capacity simultaneously,” remarked Chang Xu from the University of Sydney’s Centre for Artificial Intelligence and the Zero Emissions Institute back in 2024. “It possesses about 100 billion neurons, which it selectively recruits from different hemispheres to perform distinct tasks or modes of thought.”
Using just a few dozen watts, the brain powers the entire body; it can generate internal monologue and visual constructs from pure thought; it can recognize the shape of an incoming ball, execute complex calculations to predict its landing spot, and move a limb to intercept it—all within mere milliseconds.
This truly remarkable efficiency has long eluded replication in artificial systems, largely due to its inherent structural organization.
“A key difference between a computer and the brain lies in how each system processes information,” explained Professor Likun Luo in the 2018 publication, “The Analytical Center: Forty Scientists Explore the Biological Roots of Human Experience.”
“Computerized tasks are largely executed serially,” Luo wrote. “This is evident in how engineers program computers, establishing a sequential stream of instructions. Such a cascading, sequential flow of operations demands high precision at every juncture, as inaccuracies propagate and magnify in subsequent stages.”
The brain also employs sequential steps for specific operations, but unlike computers, this isn’t its exclusive mechanism. “Your brain also utilizes massively parallel processing,” Luo continues, “engaging a vast number of neurons and the extensive network of connections each neuron forms.”
Consider catching a ball. A computer must first detect the ball, then calculate its trajectory, estimate its final position, before finally instructing the arm to move. The brain, conversely, accomplishes all these steps almost concurrently. “By the time signals propagate through two or three synaptic connections in the retina’s photoreceptor cells, information about the ball’s location, trajectory, and velocity has been extracted by parallel neural circuits and conveyed in parallel to the brain,” Luo elaborated. “Similarly, the motor cortex (the part of the cerebral cortex governing voluntary movement) issues commands in parallel to coordinate muscle contractions in the legs, torso, arms, and wrists, ensuring the body and hands are positioned correctly—simultaneously—to receive the incoming object.”
For software engineers, this presents a compelling prospect. Could we design computing systems inspired by this massively parallel architecture? And if so, what advantages would such designs confer?
No matter how human-like contemporary AI programs may appear, they are fundamentally different at the most basic level. It’s not just that they cannot process information like us—the reverse is also true. “Traditional AI models rely heavily on backpropagation,” explained Sui Yi from the Texas A&M Engineering College, “a mechanism used to refine neural networks during training that is biologically implausible.”
Thus, achieving a computer more analogous to the human brain would necessitate a complete overhaul. New algorithms running on novel topologies would be mandatory; it would require reimagining from the ground up how connections and processes should be implemented and prioritized.
Many researchers are already pursuing this. “We have been working on removing the biological implausibility inherent in current machine learning algorithms,” Yi stated. “Our team is exploring mechanisms like Hebbian learning and Spike-Timing-Dependent Plasticity—processes that enable neurons to strengthen connections in a manner that mirrors real brain learning.”
For instance, researchers at the University of Surrey are pioneering a technique called Topographical Sparse Mapping. This method dictates that neurons connect only to those in their immediate vicinity, drastically cutting the network’s energy consumption. An additional tool, known as Enhanced Topographical Sparse Mapping, can further refine this by pruning unnecessary connections—much like how our own brains optimize their neural wiring through learning.
“By mirroring the topographical structure [of the brain], we can train AI systems that learn faster, require less power, and maintain the same accuracy,” explained Mohsen Kamelian Rad from the University of Surrey. “This is a novel way of conceptualizing neural networks, founded on the same biological principles that make natural intelligence so efficient.”
Meanwhile, teams at the University of Sydney and the University at Buffalo are developing similar neural models inspired by brain function. The objective extends beyond maximizing efficiency—though that remains a priority—to creating AI that is more intuitive, capable of leveraging fuzzy or incomplete data, or processing queries non-linearly.
In summary, “next-generation computers will look quite unlike those of the past,” Madhavan concluded. “As the volume and character of the data we gather evolve, the demands on our computational systems must change accordingly.”
“The hardware that will power the computational applications of the future must minimize its environmental footprint and be inherently green,” he cautioned. “Thanks to recent breakthroughs in neuroscience, forthcoming computers will be able to capitalize on newly unveiled biological secrets to meet the growing demand for energy-efficient computation.”
About the Author
