China is creating light chips for universal AI, smarter than humans

Scientists in China have developed a tiny modular chip that is powered by light, not electricity. Such microcircuits will be able to combine into larger structures, forming chiplets, which they want to use to train and support the operation of a future model of artificial general intelligence (AGI).

A new chip named Taichi. Forming a large puzzle of similar microcircuits, it creates a complex and powerful computing system. If properly scaled, it will be powerful enough to train and run AI in the future.

Details

AGI – it is, for now, a hypothetical advanced form of artificial intelligence that will be as intelligent as a human in terms of its cognitive abilities. AGI can be applied in many disciplines, while today's AI systems have a very narrow scope.

Some experts believe that such systems are still many years away, with the main obstacle being a bottleneck in computing power, while others believe that we will have an AI agent as early as 2027. The first steps towards this have already been taken.

In recent years scientists have begun to reach the limits of the capabilities of ordinary electronic components, especially given the rise of AI and the vast amounts of energy required to service these increasingly demanding systems. Graphics processing units (GPUs) have become key components for training artificial intelligence systems because they are better at performing parallel computations than central processing units (CPUs). But the required level of power consumption becomes unsustainable as systems get larger, the scientists say.

Light-based components could be one way to overcome the limitations of traditional electronics, including energy efficiency issues. So it is not surprising that China has taken up these studies.

What are photonic chips and how do they work

In February, scientists presented the project of a new type of photonic microchip that uses photons, or particles of light, instead of electrons to operate transistors — tiny electrical switches that turn on or off when voltage is applied. In general, the more transistors in a chip, the more computing power it has and the more energy it needs to operate. LED-based chips consume much less power and can perform calculations much faster than traditional chips because they can perform calculations in parallel.

Current photonic chip architectures for AI models consist of hundreds and thousands of parameters or training variables. This makes them quite powerful for basic tasks like pattern recognition, but large language models (LLMs) like ChatGPT are trained using billions or even trillions of parameters. An AGI agent would probably need many orders of magnitude more . Today, there are no ready-made schemes for building an AGI system.

In the new study, scientists designed Taichi to work in the same way as other light-based chips, but to scale much better than competing developments. This is because it combines several advantages of existing photonic chips — including “optical diffraction and interference”, which are ways of manipulating light in a component.

To test the design, the researchers stitched together several Taichi chips and compared their architecture to other light-based chips in key areas.

• Their architecture reached the scale of a network of 13.96 million artificial neurons — compared to 1.47 million in the next largest competing project — with a power efficiency figure of 160.82 trillion operations per watt (TOPS/W).
• The next best result they highlighted in their paper came from a study published in 2022 in which the photonic chip achieved 2.9 TOPS/W.
• Many conventional neural processors (NPUs) and other chips achieve well below 10 TOPS/W.

The researchers also claim that their Taichi-based architecture is twice as powerful as other photonic systems, but they do not specify direct test results. In tests, they used Taichi's distributed network to perform tasks including image categorization and classification, as well as generating graphical content, as a proof of concept rather than a performance comparison.

Taichi points out on the great potential of photonic computing on a chip to handle a variety of complex tasks with large network models, enabling the application of optical computing in real life, – say the engineers.

They expect Taichi to “accelerate the development of more powerful optical solutions as a critical support for the base model and the new era of AGI.”

If the technology proves viable, China could get in their hands a serious advantage over Western developments in the field of AI.