The product and its performance are detailed in two recent publications in Nature Electronics (here and here) and represents a major step toward achieving high-performance AI computing while significantly reducing energy consumption, the company said on its website.
Conventional digital computing methods for executing deep neural networks have limitations related to performance and energy efficiency. These digital systems require constant data transfers between memory and processing units, which can slow down computations and reduce energy efficiency.
To address these challenges, IBM Research has adopted the principles of analog AI, which mimics the functioning of neural networks in biological brains.
This approach involves the use of nanoscale resistive memory devices, specifically phase-change memory (PCM), to store synaptic weights.
PCM devices change their conductance through electrical pulses, allowing for a wide range of values for synaptic weights. This analog approach reduces the need for excessive data transfer by performing computations directly in the memory, resulting in improved efficiency.
The newly unveiled chip is an advanced analog AI solution featuring 64 analog in-memory compute cores. Each core incorporates a crossbar array of synaptic unit cells along with compact analog-to-digital converters, seamlessly transitioning between analog and digital domains.
Additionally, digital processing units within each core handle nonlinear neuronal activation functions and scaling operations. The chip also includes a global digital processing unit and digital communication pathways for interconnectivity.
The research team demonstrated the chip's capabilities by achieving an impressive 92.81 percent accuracy on the CIFAR-10 image dataset, setting a new standard for precision among analog AI chips.
The chip's compute efficiency, measured in Giga-operations per second (GOPS) per unit area, surpassed that of previous in-memory computing chips.
IBM says its innovation represents a significant milestone in AI hardware and can be applied across a wide range of practicalities, without revealing any examples. The chip was fabricated in IBM’s Albany NanoTech Complex but the company has not disclosed the materials used to make it.
IMF, which stands for the International Business Machines Corporation, is an American multinational technology corporation headquartered in New York and is present in over 175 countries. It manufactures computer hardware, middleware, and software, and provides hosting and consulting services in areas ranging from mainframe computers to nanotechnology.
Feel free to support our small office: @businessnewsservice (PayPal) or IBAN - RO50BTRLEURCRT0490900501, Swift - BTRLRO22, Beneficiary - Rudeana SRL.
Not feeling like donating? Then click on banners on our website to generate ad revenue. Any help is welcome.