Will brain-like computers become the "darling" of the AI ​​era?

The power demands of modern computers are growing at an alarming rate, and many technology companies are working hard to develop more energy-efficient hardware. But is it possible to build a computer with a completely new architecture that could achieve a quantum leap in energy efficiency?

Akida neuromorphic chip launched by BrainChip.
Image source: BrainChip official website

Some companies have given a positive answer. They are using neuromorphic computing technology to create computers that think like the brain, that is, "brain-like computers." This innovative technology aims to imitate the way the human brain processes information, and may set off a revolution in many fields such as artificial intelligence (AI) and robotics.

The human brain may be the ultimate form of computer

The human brain runs on very low energy consumption, yet it can accomplish complex and diverse tasks. According to the U.S. Physics Today website, the human brain runs on about 12-20 watts of power, which accounts for 20% of the body's metabolic rate; in comparison, some desktop computers run on about 175 watts, and cutting-edge AI accelerators such as Nvidia's H100 run on 300-700 watts of power.

In neuromorphic computing, electronic devices mimic neurons and synapses and are connected to each other in a way that resembles the brain's networks.

Neuromorphic computers have some fundamental features in their design that set them apart from traditional computers. First, neuromorphic computers do not have separate memory and processing units. These tasks can be performed together at the location of each neuron on the chip, so there is no need to transfer data between memory and processors, reducing energy consumption and speeding up processing.

Second, whereas in traditional computing, every part of the system is always on and able to communicate with any other part, the simulated neurons and synapses in a neuromorphic system are activated only when needed, saving energy.

In addition, modern computers are digital and use 1 or 0 to represent data; the electrical signals in the brain are not simply composed of 0 and 1, and neuromorphic computers can also simulate this aspect of the brain.

Hardware and software are the dual pillars

Neuromorphic computing relies on two fundamental technology pillars: hardware and software.

In terms of hardware, scientists are developing specific neuromorphic chips. For example, Intel released the prototype neuromorphic chip Loihi 2 in 2021. This chip has an area of ​​31mm2 and its processor can package up to 1 million artificial neurons.

In April this year, Intel announced the creation of Hala Point, the world's largest neuromorphic system, designed to support future brain-like AI research. The system is built on the Loihi 2 processor, with up to 1.15 billion neurons and 128 billion synapses, and is up to 200 times faster than the human brain. Intel said that Hala Point's neuron capacity is roughly equivalent to that of an owl's brain, making it the world's largest neuromorphic computer to date.

In addition to Intel, IBM also launched its latest brain-like chip prototype "North Pole" last year. This chip is an upgraded version of the previous "True North" chip. Tests show that it is more energy-efficient, space-saving and faster than other chips on the market. Currently, the research team is working to combine these chips into larger-scale systems.

In addition, smaller neuromorphic companies such as Australian AI chip maker BrainChip, Chinese AI chip startup SynSense, and Dutch neuromorphic processor company Innatera are also actively investing in R&D in this field.

On the software side, algorithms and computational models are being developed that mimic the way the brain learns and processes information, such as artificial neural networks and deep learning.

Prospects for commercial applications are optimistic

According to the BBC, the future commercial applications of neuromorphic computers are mainly divided into two major areas: one is to provide a more energy-efficient and higher-performance platform for AI applications, including image and video analysis, speech recognition, and large language models that power chatbots such as ChatGPT; the other is "edge computing", that is, real-time data processing on networked devices. Self-driving cars, robots, mobile phones, wearable technology, etc. can greatly improve efficiency through "edge computing" applications.

However, technical challenges still exist. One of the main obstacles to the development of neuromorphic computing is the development of software to adapt to the operating requirements of these unique hardware. Although the hardware has gradually matured, how to activate its potential in a completely new programming way remains an urgent problem to be solved. In addition, cost is also a major challenge. Whether it is silicon-based or other materials, it requires high costs to manufacture new neuromorphic chips.

In May this year, German technology company SpiNNcloud Systems announced that they are developing a neuromorphic supercomputer called SpiNNaker2 that can simulate at least 10 billion neurons and plans to commercialize it. This hybrid AI high-performance computer system based on the principles of the human brain has undoubtedly injected new vitality into the field of neuromorphic computing.

Tony Kenyon, a neuromorphic research expert at University College London, said: "Although there is no 'killer-level' application yet, neuromorphic computing will significantly improve energy efficiency and performance in many areas. As this technology matures, we will see its widespread application."