ACIM(Analog Computing-in-memory) is one of the most promising non-von Neumann methods that can generate energy-efficient deep learning inference hardware. However, MAC(Multiply and Accumulation) is executed in the analog section, and the weight variation stored in analog synapses may cause errors that may impact the accuracy of computing.

 

In conducting the 4-8bit operation required by the edge system for image inference, we focused on identifying the overlap (variation) limitation for each weight level of the cell and the linearity of weight level values in order to identify the process control factors to make improvement on linearity and variation and quantify its effectiveness in enhancing the recognition rate.

 

In the IMW 2022 research paper, 1) we have secured the actual wafer data on 16-weight level by establishing a baseline for material, process and structure of analog synapse device for ACiM, and 2) checked the possibility of matrix computing. 3) We have applied process variations and analyzed contributions based on a compact model, of which consistency rate is above 97% compared to the actual measurement data, and have conducted various experiment evaluations on thickness, composition, size, crystalline structure, roughness, etc., to verify the optimization method. While other process parameters showed less than 2% of improvement impact on variation, roughness was the most effective parameter in improving variation, as it can induce the concentration of the electric field to create various levels of conductive filaments. 4) We have conducted a CIFAR10, VGG8 model-based evaluation utilizing an adjusted architecture simulator to reflect the wafer data on each level, and checked that the recognition rate was improved from 35.77% to 84.35%.

 

This result can provide a guideline on the design and optimization of the resistive synapse device to realize ACiM, and will be used to identify the key control weight level and uniformity management method by position in 16 weight level based on uniformity analysis on each weight level in the future. This memory-centric NVM-based ACiM technology is expected to contribute to RTC’s preparation for beyond memory technology.

The publication: Link