Scientists in China have designed a small, modular chip that runs on light instead of electricity – and they want to use it to train and operate future artificial general intelligence (AGI) models.
The new chiplet, called a “Taichi,” is a small piece of a larger jigsaw made up of many individual chiplets (including Taichi modules) that together can form a sophisticated and powerful computing system. If scaled up enough, it would be powerful enough to train and operate AGI in the future, the scientists argued in their paper, published April 11 in the journal. science.
AGI is a hypothetical advanced form of artificial intelligence (AI) that will theoretically be as smart as humans in terms of its cognitive reasoning abilities. AGI can be applied in many fields, whereas today’s AI systems can only be applied in a few.
Some experts believe that such systems are many years away, with a bottleneck in computing power being a key blocker, while others believe that we will have an AGI agent as soon as 2027.
In recent years, scientists have begun to push the limits of traditional electronics-based components, especially given the advances in AI and the sheer amount of power required to service these increasingly demanding systems.
Related: A computer chip powered by light can train AI faster than components powered by electricity.
Graphics processing units (GPUs) have emerged as key components in training AI systems, as they are better at performing parallel computations than central processing units (CPUs). But scientists argue that required levels of energy consumption are becoming unsustainable as systems grow larger.
Light-based components can be a way to overcome the limitations of traditional electronics – including energy efficiency issues.
Searching for light for supernatural AI
Scientists have previously outlined a design for one. A new type of photonic microchip in February, which uses photons, or particles of light, rather than electrons to operate transistors – tiny electrical switches that turn on or off when a voltage is applied. In general, the more transistors a chip has, the more computing power it has and the more power it needs to operate. Light-based chips consume much less energy and can perform calculations much faster than conventional chips, because they can perform calculations in parallel.
Current photonic chip architectures for AI models contain hundreds or thousands of parameters, or training variables. This makes them powerful enough for basic tasks like pattern recognition, but large language models (LLMs) like ChatGPT are trained using billions or trillions of parameters.
An AGI agent would likely require orders of magnitude more—as part of a broader network of AI architectures. Today, blueprints for building AGI systems do not exist.
In the new research, the scientists designed Tychi to work like other light-based chips, but scale much better than competing designs, they said in their paper. That’s because it combines several advantages of existing photonic chips — including “optical diffraction and interference,” which are ways to couple light into a component.
To test the design, the researchers assembled several Taichi chiplets and compared their architecture with other light-based chips in key areas.
Their architecture achieved a network scale of 13.96 million artificial neurons – compared to 1.47 million in the next largest competing design – with an energy efficiency metric of 160.82 trillion operations per watt (TOPS/W). The next best result they highlighted in their paper came from The research was published in 2022., in which a photonic chip achieved 2.9 TOPS/W. Many get traditional neural processing units (NPUs) and other chips. Less than 10 TOPS/W.
The researchers also claimed that their Taichi-based architecture is twice as powerful as other photonic systems, but did not directly cite them. In tests, meanwhile, they used the distributed Taichi network as a proof of concept rather than benchmark performance, to perform tasks including image classification and classification as well as image content creation.
“Taichi demonstrates the great potential of on-chip photonic computing to process multiple complex tasks with large network models, enabling real-life applications of optical computing,” the scientists said. We expect Taichi to accelerate the development of foundational models and more powerful optical solutions as critical support for a new era of AGI.