The new AI chip, which operates like a human brain, has 64 AIMC cores for deep neural network computational tasks. (Image Credit: IBM)
IBM recently revealed its prototype mixed-signal analog AI chip that operates like a human brain and runs various complex deep neural network computational tasks. It has the ability to boost AI efficiency and decrease the battery power consumption of a computer and smartphone.
The cutting-edge chip features 64 analog in-memory cores, each with a 256x256 crossbar array of synaptic unit cells. "It also implements the digital activation functions and additional processing involved in individual convolutional layers and long short-term memory units," IBM wrote in the paper.
Every core (tile) comes with compact, time-based analog-to-digital converters for analog and digital switching. In addition, they have lightweight digital processing units to run simple nonlinear neuronal activation functions and scaling operations. The center of the chip contains a global processing unit that supports complex operations essential for running certain kinds of neural networks. IBM interconnected the cores with the global digital processing unit, similar to how neurons interconnect in the human brain.
"Each tile can perform the computations associated with a layer of a DNN model. The synaptic weights are encoded as analog conductance values of the PCM devices," IBM wrote in the blog post.
The IBM team also discovered that their AI chip ran with an accuracy of 92.81% on the CIFAR-10 image dataset. They say this is the most accurate chip compared to others that have similar technology. Also, the chip's "measured throughput per area for.8-bit input-output matrix multiplications of 400 GOPS/mm2 is over 15 times higher than multi-core, in-memory computing chips" with comparable energy efficiency.
This chip could eventually replace those that run heavy AI applications in smartphones and computers. Its energy efficiency means that it has the potential to be powered in low-power or battery-constrained environments, including cars, mobile phones, and cameras. Cloud providers can even use them to shrink energy costs and their carbon footprint.
Have a story tip? Message me at: http://twitter.com/Cabe_Atwell