IBM’s analogue AI chip achieves 400Gop/s/mm2

Author: EIS Release Date: Aug 21, 2023

IBM has revealed an analogue in-memory IC for implementing neural networks.

 

IBM proposed analogue AI chip

The device stores weights locally as analogue levels as conductance in phase-change memory, and implements analogue multiply-accumulate calculation.

 

For analogue AI processing “two key challenges need to be overcome: These memory arrays need to compute with precision on par with existing digital systems, and they need to be able to interface seamlessly with digital compute units, as well as digital communication fabric on the chip”, said the organisation.

 

 

Fabricated in-house with back-end phase-change memory, the 14nm cmos IC is composed of 64 analogue in-memory computing tiles, each with a 256 x 256 crossbar array of synaptic unit cells.

 

 

ADCs on each tile interface with the digital side of the IC. “Each tile is also integrated with lightweight digital processing units that perform simple nonlinear neuronal activation functions and scaling operations,” according to IBM, and “a global digital processing unit is integrated in the middle of the chip that implements more complex operations that are critical for the execution of certain types of neural networks”.

 

Each tile can be used to implement a layer of a DNN (distributed neural network) model, and weight accuracy was equvilent to 3bit or 4bit.

 

Using the IC, an accuracy of 92.81% was demonstrated on the CIFAR-10 image dataset.

 

“We believe this to be the highest level of accuracy of any currently reported chips using similar technology,” said IBM.

 

8bit input-output matrix vector multiplication density was 400Gop/s/mm2, and it achieved peaks of 63Top/s and 9.76Top/W.

 

Nature Electronics will be describing the work in the paper: ‘A 64-bit mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference’.