LLMs: “Biological neurons consume a thousand times less energy than Nvidia chips”

• Algorithm
• Deep learning
• Machine learning
• Prospective
• Technology

Share on Twitter : LLMs: “Biological neurons consume a thousand times less energy than Nvidia chips” (New window) Share on Linkedin : LLMs: “Biological neurons consume a thousand times less energy than Nvidia chips” (New window)

How did you come up with the idea of using biological neurons rather than continuing with traditional computing for AI?

Initially, we began working on simulating very specific neuron models called “spiking neurons.” These models are much more sophisticated and realistic in their simulation than those currently used by standard AI systems, such as ChatGPT. We very quickly hit a technical wall because the colossal computing power required to simulate such networks is a major obstacle, and training them is extremely complex to achieve numerically. We crossed paths with neurobiologists who were studying human neurons cultured in vitro. Rather than trying to calculate and simulate these complex networks numerically, we could directly use the natural transfer function of the biological network.

In practical terms, how do you interact with these human cells in culture to transmit information to them?

Communication with the cells occurs primarily through electrodes used either to send electrical impulses directly into the tissue or to receive and measure the action potentials emitted by the neurons. We can also release neuromodulators, such as serotonin, or neurotransmitters, such as glutamate, directly into the culture medium. This allows us to influence the overall behavior of the neurons and modify their state. In fact, if you visit our website, you can observe the real-time activity of our hundreds of thousands of neurons, which is streamed 24 hours a day, 7 days a week.

One of the major challenges in AI is training. How do we teach a biological network to perform a specific task without being able to program its code?

The mechanism is very similar to what we’re doing right now as we talk. When I speak to you, I send impulses that reach your nervous system, which gradually updates your connections, and you respond to me through your own nerve impulses. In artificial intelligence, we use algorithms designed thirty years ago, such as backpropagation, to mathematically update synaptic connections and train the network. In our case, with living cells, we have no direct access to synaptic connections. However, nature proves to us that a nervous system is capable of learning. Currently, a large part of basic research is focused precisely on figuring out how to properly stimulate neurons so that their output meets expectations.

The stated goal is to replace servers equipped with Nvidia GPU chips for environmental reasons. What is the actual benefit in terms of energy consumption?

The long-term goal is, in fact, to replace current digital servers. If we consider energy efficiency at the cellular level alone, a biological neuron consumes about one million times less energy than an artificial neuron. But the reality is even more complex: today, to accurately simulate a single biological neuron in all its complexity, it can sometimes take tens of thousands of artificial neurons. Taking everything into account—and considering that biological systems operate more slowly—we estimate that our systems will consume at least 1,000 times less energy than a conventional system. I am convinced that this will eventually render digital systems obsolete for specific applications such as generative AI.

Are there inherent limitations to this biological approach compared to silicon? Will we be able to do everything with your bioprocessors?

There are very clear limitations, particularly in terms of speed. Biological neurons are slow—about 10 to 500 times slower than their electronic counterparts. Consequently, you’ll never be able to process a million images per second with a biological network to perform high-speed image recognition, a field in which artificial networks excel. However, when it comes to use cases involving relatively low data throughput but requiring a higher “processing rate,” biological networks are ideal.

Where do you stand in terms of commercialization and research? Who is using FinalSpark today?

Currently, the technology is in a remote experimental phase available to researchers. We lease this access for between $1,000 and $5,000 per month to laboratories around the world, such as the Université Côte d’Azur in France. Their research is very diverse: some are trying to understand how a brain organoid can encode tactile data, while others are using it to study “connecthome”—that is, trying to deduce how neurons connect to one another based on measured electrical impulses. The enthusiasm extends beyond the academic sphere: the U.S. Defense Advanced Research Projects Agency (DARPA) recently allocated $35 million to this type of biocomputing project—for example, to teach biological networks to play Pac-Man or to equip drones capable of detecting explosives.

This text has been translated by an artificial intelligence.