BPUs Have a Mind of Their Own

Anyone with even a passing interest in artificial intelligence (AI) has undoubtedly heard plenty about GPUs and TPUs. These are the workhorses that make the execution of massive AI algorithms possible. But what about BPUs? These so-called Brain Processing Units are not exactly a mainstream technology, so it is understandable if you are unfamiliar with them. They may not lurk in the shadows too much longer, however, as they have the potential to revolutionize the entire field.

Many AI algorithms — such as artificial neural networks — are designed to very roughly approximate some aspect of the human brain’s functions. However, traditional hardware accelerators do not function very much at all like a brain. This mismatch is at least part of the reason why the data centers running these algorithms may consume as much energy as a small city while our brains only need about as much energy as a light bulb. With some refinement, BPUs could offer the efficiency that is needed to practically run much larger, more powerful AI algorithms in the future.

A pioneering project exploring BPUs is currently underway as a collaboration between Daito Manabe, the SoftBank Research Institute of Advanced Technology, and the Ikeuchi Laboratory at the University of Tokyo. Their research focuses on the use of cerebral organoids — tiny clusters of artificial brain tissue cultivated from induced pluripotent stem cells — as the core of a new computing paradigm. These miniature brain structures, which can contain up to 100 million neurons, are being used in conjunction with custom electronics to both stimulate and analyze their neural activity.

Recently, an exhibition at the University of Tokyo showcased the current state of this research through three experimental demonstrations, each highlighting a different aspect of how biological neurons could be integrated into computing and AI applications.

One of the most interesting experiments explored how cerebral organoids process music. In this study, researchers delivered different musical genres — including techno, classical, and ambient noise — to the neural tissue using optogenetic stimulation. By analyzing how the organoids responded to these stimuli, scientists sought to uncover fundamental principles of music perception in biological systems. The results were compared against baseline responses to non-musical stimuli, such as white noise and a pure sine wave. This experiment provided insights into how neural networks, even in their simplest form, react to and differentiate auditory patterns.

Another experiment demonstrated the potential of cerebral organoids to serve as control units for autonomous robots. In this setup, a quadrupedal robot was equipped with a neural processing system based on cultivated brain cells. A ceiling-mounted camera tracked the robot’s movements and transmitted real-time data to the organoids as electrical signals. The system used a reward-and-punishment approach, where electrical stimulation was associated with free space, while its reduction indicated obstacles. Over time, the organoid-based system autonomously developed obstacle-avoidance behaviors, demonstrating a biological learning mechanism that could one day complement or even replace traditional AI-based robotic control systems.

The third demonstration examined how cerebral organoids process rhythm and whether they can generate rhythmic patterns themselves. By stimulating the organoids with periodic electrical pulses and later recording their spontaneous neural activity, researchers observed distinctive rhythmic responses. They even introduced a feedback loop where the organoids’ own activity was converted into MIDI drum patterns and reintroduced as stimulation. This setup mimicked a recurrent neural network but with the added complexity and unpredictability of biological neurons. Researchers believe this could provide clues about how humans may have first developed the ability to recognize their own speech.

While still in its early stages, this research represents a meaningful step toward integrating biological intelligence with computing. Unlike traditional AI, which relies on artificial neural networks running on power-hungry processors, BPUs harness the natural efficiency of biological neurons. If refined, they could lead to more energy-efficient AI systems, advanced robotics, and even new forms of human-computer interaction.