In an era marked by soaring energy consumption in data centers and an insatiable demand for advanced chips, a novel approach is emerging: the use of biological brain cells. Cortical Labs, an Australian startup, has unveiled plans to construct two “biological” data centers in Melbourne and Singapore. These facilities will be equipped with specialized neuron-filled chips, the same technology that has already demonstrated the ability to play games like Pong and Doom.
Cortical Labs is among a select group of organizations pioneering biological computing. Their systems integrate neuronal cells with microelectrode arrays. These arrays can both stimulate the cells and meticulously measure their responses when presented with data, effectively creating a living computer. Notably, the company recently showcased its flagship product, the CL1, learning to play the video game Doom within a single week.
The newly announced data centers represent a significant step forward. The Melbourne facility is slated to house approximately 120 CL1 units. The second, developed in partnership with the National University of Singapore, will initially contain 20 CL1s. However, Cortical Labs envisions expanding this to 1,000 units in a larger center, pending regulatory approval. This expansion aims to bolster their cloud-based brain-computing services.
Michael Barros from the University of Essex in the UK acknowledges the global effort in developing biological computers like the CL1. He points out the inherent difficulty and user accessibility challenges often associated with these systems. “We spend a lot of money and sweat to build these [systems],” Barros remarked. He believes Cortical Labs’ initiative to make biocomputers accessible on a large scale is unprecedented. Barros, who already utilizes Cortical Labs’ cloud services for his own research, stated, “They’ll be the first ones to do that.”
While these systems can indeed be trained for tasks as straightforward as playing Doom, the precise mechanisms of neuronal operation and optimal training methods for tasks like machine learning remain areas of active investigation, according to Reinhold Scherer, also at the University of Essex. “Having access to this allows you to explore learning, training and programming,” he explained. “You don’t program neurons like standard computers.”
Cortical Labs asserts that its biological data centers will operate with substantially less power than conventional computing setups. They estimate that each CL1 unit requires around 30 watts, a stark contrast to the thousands of watts demanded by state-of-the-art silicon-based AI chips.
Paul Roach of Loughborough University in the UK commented on the potential for significant energy savings as these systems scale up. “When we scale up and have these as whole rooms, just like you have now with data servers, then there could be huge power savings,” Roach noted. He also highlighted that while biological data centers might require additional resources, such as nutrients for the neuronal chips, they should necessitate far less cooling compared to traditional computing infrastructure. “The amount of energy that’s saved with [Cortical Labs’s] figures is fairly conservative,” he added.
Despite these promising developments, the technology is still in its nascent stages, according to Tjeerd olde Scheper at Oxford Brookes University. Having worked with a competitor, FinalSpark, Scheper offered a cautious perspective. “Is it going to work as people might think? No, we’re still in the early days of this development.”
Direct comparisons in terms of physical size are complicated because CL1 chips cannot perform conventional calculations like standard silicon AI chips. However, the proposed biological data center, with its hundreds of biological chips, would be significantly different from the hundreds of thousands of graphics processing units (GPUs) found in the largest AI data centers currently in operation.
Steve Furber from the University of Manchester expressed a similar sentiment regarding the technology’s readiness. “I think it is a very long way from production-ready. It’s a very big step from a small network playing a computer game to a large language model,” he stated.
Key challenges persist, particularly in understanding how to store the results of neuronal training in a retrievable memory format. Furthermore, running actual computational algorithms on these biological systems, beyond specific training tasks like playing video games, is another area requiring further research. Scherer also pointed out the issue of retraining neurons after they have completed a given task. “Whatever they are trained on is lost when the culture ends its life, so there needs to be a proper retraining,” he explained. “Then it’s not an optimal solution to keep a technology going if you need to retrain every 30 days.”