Brain Cells Powering Data Centers: The Dawn of Biological Computing

The insatiable demand for energy and advanced chips in traditional data centers has sparked a search for alternative computing solutions. Australia-based startup Cortical Labs is stepping into this arena by announcing the construction of two novel “biological” data centers in Melbourne and Singapore. These facilities will house chips populated with biological neurons, a technology the company has previously demonstrated capable of playing classic video games like Pong and Doom.

Cortical Labs stands among a select group of companies pioneering the development of biological computers. These systems integrate neuronal cells with microelectrode arrays. This unique architecture allows for the stimulation and measurement of cellular responses when presented with data. Earlier in the month, the company showcased the learning capabilities of its flagship computer, the CL1, which successfully learned to play the game Doom within a single week.

The company has now unveiled plans for two distinct data centers. The Melbourne facility is slated to feature approximately 120 CL1 units. The second center, a collaboration with the National University of Singapore, will initially house 20 CL1 units. Cortical Labs aims to expand this to 1000 units in a larger, purpose-built data center, pending regulatory approval. This expansion is intended to bolster their cloud-based brain computing service offerings.

Michael Barros of the University of Essex notes that while biological computers like the CL1 are being developed globally, they present significant construction challenges and lack widespread accessibility. “We spend a lot of money and sweat to build these [systems],” he commented. Barros, who already utilizes Cortical Labs’ cloud services for his research, believes their initiative is groundbreaking. “What [Cortical Labs] is doing is essentially allowing its biocomputer to be accessible at a large scale,” he stated, adding, “They’ll be the first ones to do that.”

While these systems exhibit proficiency in relatively straightforward tasks, such as playing Doom, the precise mechanisms of neuronal function and optimal training methodologies for tasks like machine learning remain subjects of ongoing investigation. Reinhold Scherer, also affiliated with the University of Essex, highlighted the potential of accessible biological computing. “Having access to this allows you to explore learning, training and programming,” Scherer explained. “You don’t program neurons like standard computers.”

Cortical Labs asserts that its biological data centers will consume considerably less energy than conventional computing systems. The company claims each CL1 unit requires approximately 30W, a stark contrast to the thousands of watts needed by state-of-the-art conventional AI chips. Paul Roach from Loughborough University suggested that scaling up these systems into entire rooms, mirroring current data server infrastructure, could yield substantial power savings. He also pointed out that biological data centers might require different resources, such as nutrients for the neuronal chips, but would likely need significantly less cooling than traditional setups. Roach believes the energy savings projected by Cortical Labs are conservative.

However, Tjeerd olde Scheper from Oxford Brookes University, who has experience with a competing biological computing company, FinalSpark, cautioned that the technology is still in its nascent stages. “Is it going to work as people might think? No, we’re still in the early days of this development.”

Direct size comparisons are challenging because CL1 chips cannot perform conventional calculations like standard silicon-based AI chips. Nevertheless, the proposed biological data center plans incorporate hundreds of biological chips, a number significantly lower than the hundreds of thousands of graphics processing units (GPUs) found in the largest AI data centers.

Steve Furber of the University of Manchester expressed reservations about the technology’s readiness for widespread production. “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 an LLM,” he remarked.

Key challenges persist, including the unclear methods for storing training results from neurons in a persistent memory format or for executing general computational algorithms beyond task-specific training, such as for video games. Furthermore, retraining neurons after a particular task is completed presents another hurdle. Scherer noted, “Whatever they are trained on is lost when the culture ends its life, so there needs to be a proper retraining. Then it’s not an optimal solution to keep a technology going if you need to retrain every 30 days.”