Researchers at IBM have developed a new hardware neuromorphic chip that is highly efficient and low power. In a paper published in Nature Electronics, the scientists describe a 64-core mixed-signal memory computing chip that utilizes a structure imitated from the human brain.
The main problem with current artificial neural networks is their demanding performance and power consumption. Most networks run on remote servers because they require powerful hardware. But IBM scientists decided to find a solution to this problem by turning to nature.
Using the principles of the human brain, the researchers developed a chip with mixed-signal memory. This chip has analog cores in memory, each containing an array of synaptic cellular units. It uses transducers to transition between analog and digital states, allowing it to efficiently perform complex computations at low power consumption.
One of the major achievements of this chip is its accuracy in image processing. According to IBM, the chip has achieved an accuracy of 92.81% on the CIFAR-10 dataset, which is widely used in machine learning tasks. This means that the mixed-signal chip has great potential for processing complex tasks in power-constrained environments such as cell phones, cars and cameras.
Thanos Vasilopoulos, one of the study’s co-authors, notes that the human brain is capable of achieving outstanding performance while consuming little power. That’s why researchers turned to the structure of the brain when developing the new chip.
This new chip opens up new possibilities for the application of artificial intelligence in various fields where limited battery power is a problem. For example, mobile devices, cars and cameras will be able to use artificial intelligence to perform complex tasks without significant power consumption.
Researchers at IBM continue to work on improving this chip and its application in various fields. They hope that their development will help create more efficient and energy-efficient artificial intelligence systems.