mythofx

[mythofx]

The Emergence of Neuromorphic Hardware for AI Processing

Until recently, AI systems used massive amounts of energy to run their algorithms. But the rise of neuromorphic hardware—which mimics how the brain processes data and performs its operations—may make it possible to process artificial intelligence algorithms on a chip that consumes less power and uses less memory.

But bringing AI on board with these devices won’t be without its challenges. It will take significant effort to adapt current AI software to work with neuromorphic chips. And even after that work is completed, the underlying architecture of these systems may still be difficult to understand.

The hardware of neuromorphic chips combines elements of both digital and analog circuitry in a way that closely mirrors how neurons behave. Each analog neuron in a neuromorphic chip stores information as changes in its electrical resistance or conductance rather than as binary bits. This feature makes the chips closer to the physics of real neurons than existing silicon complementary metal oxide semiconductor (CMOS) processors. The analog neurons also operate at a faster speed than their digital counterparts, making them more effective for executing some types of algorithms.

Another key aspect of neuromorphic hardware is that it operates at a lower voltage than existing silicon chips. This allows the analog neurons to handle a much higher load of electrical signals, which could improve processing performance. Additionally, the low voltage reduces the amount of heat that the chips generate, a critical factor in extending their lifespans.

Neuromorphic chip developers are also working to emulate the plasticity that occurs in biological neural networks, a process known as adaptive technology viewer website learning. This is important because it’s what enables the brain to learn and adapt to its environment, which in turn makes it more effective at performing complex tasks. To accomplish this task, the researchers are using memristive devices, which act like switches with variable resistance or conductance and can record a history of states. These devices can be programmed to act like the synapses that send neurons signals by controlling their conductance based on previous programming (top-down co-design).

A neuromorphic computer system developed by Intel Labs, the research division of the CPU-centric company, utilizes 130,000 Loihi chips to model and simulate a small-scale grey matter neural structure with eight million artificial neurons. Its ability to process large amounts of information in parallel is one of its most promising features.

Intel and other hardware manufacturers are working to develop neuromorphic hardware that can operate on the edge of a network. This would allow the AI algorithms to be executed locally on a machine or vehicle rather than in a cloud computing facility or offsite data center. The technology is particularly useful for applications that require real-time data processing, such as autonomous vehicles or advanced sensors. The University of Michigan’s memristor array chip, which was announced this week, is a good example. Its capabilities suggest that localized AI may be a reality sooner than many might expect.