Jane Edwards
IBM’s research arm has developed an analog artificial intelligence-based chip designed to perform deep neural network inference tasks.
The analog AI chip comes with 64 analog in-memory compute cores each containing 256-by-256 crossbar array of synaptic unit cells and digital processing units that could carry out scaling functions and nonlinear neuronal activation operations, IBM Research said Thursday.
The chip developed at IBM’s Albany NanoTech Complex in New York features a global digital processing unit meant for complex functions and digital communication pathways.
In a paper published in Nature Electronics, IBM Research scientists said they encoded the synaptic weights as analog conductance values of phase-change memory devices and found that each tile in the analog AI chip can make computations related to a layer of a DNN model.
The company described the 64-tile chip as an energy-efficient, mixed-signal architecture that can perform computing functions on the same level as current digital systems and can merge seamlessly with a digital communication fabric and other digital processing units.
Join the ExecutiveBiz Trusted Artificial Intelligence and Autonomy Forum on Sept. 12 to hear public sector leaders and technology experts as they talk about the opportunities and risks associated with generative AI and related tools. Register here.
