A ground-breaking analog AI processor from IBM Research has been disclosed, and it performs difficult computations for deep neural networks (DNNs) with astounding efficiency and precision.
This innovation is a big step toward obtaining high-performance AI computation while significantly reducing energy use. It was just published in a study in Nature Electronics.
Performance and energy efficiency are constrained by the current method of running deep neural networks on standard digital computer platforms. The continual data transfer that these digital systems require from memory to processor units slows computations and hinders energy optimization.
IBM Research has adapted the ideas of analog AI, which mimics how neural networks work in biological brains, to address these issues. In this method, phase-change memory, a type of nanoscale resistive memory device, is used to store synaptic weights.
Electrical pulses used by PCM devices to change their conductance allow for a range of synaptic weight values. As computations are carried out directly in memory, this analog method reduces the need for superfluous data transfer and improves efficiency.
A cutting-edge analog AI solution, the newly unveiled chip has 64 analog in-memory processing cores. A crossbar array of synaptic unit cells and small analog-to-digital converters are integrated into each core to enable seamless switching between the analog and digital domains. Additionally, each core’s digital processing units control scaling and nonlinear neural activation processes. The chip also features a worldwide digital processing unit and interconnected digital communication routes.
By obtaining an accuracy of 92.81 percent on the CIFAR-10 picture dataset, the research team showed the chip’s capabilities and set a record for analog AI devices.
In comparison to earlier in-memory computer chips, the throughput per area, measured in Giga-operations per second (GOPS) by area, demonstrated its greater compute efficiency. This ground-breaking chip is a milestone achievement in the field of AI hardware because to its energy-efficient design and improved performance.
The innovative design and amazing capabilities of the analog AI chip pave the groundwork for a time when a wide range of applications can access energy-efficient AI processing.
The innovation from IBM Research is a turning point that will spur improvements in AI-powered technologies for years to come.
