Chinese scientists have unveiled Taichi-II, an upgraded all-optical artificial intelligence (AI) chip that could one day be used in artificial general intelligence (AGI) systems.
The first Taichi chip was unveiled by the researchers in April 2024. Rather than relying on electronic components, the tiny, modular device is powered by photons, or particles of light. These photons power tiny on-board electrical switches that turn on and off when a voltage is applied to them.
According to a report by the South China Morning Post (SCMP), compared to its predecessor, Taichi-II is said to be 40 percent more accurate at classification tasks that classify and identify different types of information, and is “six orders of magnitude” (i.e. one million times) more energy efficient in low-light conditions.
The researchers achieved this dramatic performance boost by training an AI directly on the optical chip rather than relying on digital simulations, a process the scientists call “full forward mode.” They described their findings in a study published August 7 in the journal Nature.
Full forward mode is an AI training method where data moves in only one direction – forward. This differs from traditional training methods where data is typically processed in multiple iterative steps. As light passes through the chip, it interacts with tiny components that adjust its direction and adjust its phase and intensity. This instantly changes the parameters of the AI model, allowing it to learn in real time without repeated processing.
While the full forward mode allows photon chips to operate even faster than before, these chips already have significant advantages over conventional chips: light-based chips consume much less energy than conventional chips and can process multiple signals simultaneously, allowing them to perform calculations much faster. This is because photons, unlike electrons, can travel at the speed of light and do not generate heat as they travel through the chip, allowing for faster and more efficient processing.
Related: Startup claims 'quantum-inspired' laser computing is more effective than both supercomputing and quantum computing
Researchers previously said the Taichi chip works similarly to other light-based chips but is much more scalable than competing designs because the chiplet combines some of the advantages of existing photonic chips, such as “optical diffraction and interference,” which refers to the way light can be manipulated within a chiplet.
AI chips that operate entirely photonically could eventually lead to AGI models – extremely powerful AI systems with human-like intelligence and reasoning capabilities, and the ability to learn new skills beyond the limits of training data.
While this hypothetical technology is still years away from becoming a reality (at least according to most predictions), Taichi's developers suggested in an April paper that the chip's modular architecture could allow multiple chiplets to be combined to build extremely powerful AI systems.
They demonstrated this in experiments, stitching together several Taichi chiplets and comparing their performance in key areas with other light-based chips: the combined system was able to simulate a network of about 14 million artificial neurons, far more than the 1.47 million neurons achieved by the next best design.
This was achieved while being extremely energy efficient, performing over 160 trillion operations per watt of power consumption. By comparison, photonic chips in 2022 can only process 3 trillion operations per watt, and most conventional chips designed for similar tasks typically perform fewer than 10 trillion operations per watt.
Researchers said Taichi-II is a key step in moving light-based AI chips from theory to practical use and meeting the growing demand for high-power, low-energy computing, SCMP reported. This will be key to developing AGI models, but concerns remain about its impact.
