- August 14, 2023
- AI Projects
IBM Research has revealed an innovative analog AI chip that showcases exceptional efficiency and precision when handling intricate computations for deep neural networks (DNNs).
This remarkable achievement, outlined in a recent publication in the journal Nature Electronics, represents a substantial step forward in the pursuit of high-performance AI computing while significantly conserving power.
The conventional method of running deep neural networks on typical digital computing structures presents challenges in terms of both performance and energy efficacy. These digital systems involve continuous data transfer between memory and processing components, causing computation slowdowns and reduced energy optimization.
To address these issues, IBM Research has leveraged analog AI principles, mirroring the operation of neural networks in biological brains. This technique involves the utilization of nanoscale resistive memory devices, particularly Phase-change memory (PCM), to store synaptic weights.
PCM devices adjust their conductance via electrical impulses, allowing for a range of values for synaptic weights. This analog approach minimizes the need for excessive data transfer, as computations take place directly within the memory, leading to heightened efficiency.
The recently introduced chip stands as an advanced analog AI solution comprising 64 analog in-memory compute cores.
Each core incorporates a crossbar array of synaptic unit cells alongside compact analog-to-digital converters, enabling smooth transitions between analog and digital realms. Additionally, digital processing units within each core manage nonlinear neuronal activation functions and scaling operations. The chip also includes a global digital processing unit and digital communication pathways for interconnectedness.
The research team showcased the chip’s prowess by achieving a remarkable accuracy of 92.81 percent on the CIFAR-10 image dataset — an unparalleled level of precision for analog AI chips.
The data processing rate per unit space, quantified in terms of Giga-operations per second (GOPS) for each unit area, highlights its exceptional computational efficiency in contrast to earlier in-memory computing chips. This inventive chip’s energy-conscious design, combined with its heightened performance, positions it as a noteworthy accomplishment in the domain of AI hardware.
The distinct architecture and impressive capabilities of the analog AI chip establish a foundation for a future in which energy-efficient AI computation is feasible across a diverse array of applications.
IBM Research’s breakthrough marks a pivotal juncture that is poised to drive advancements in AI-powered technologies for the foreseeable future.
