3D NAND AI Accelerator Inference Circuit with Pre-charged Line Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inference operation of 3D NAND artificial intelligence accelerators consumes significant time and energy due to the need for calculating large amounts of data during Multiply-Accumulation (MAC) operations, leading to high power consumption.
Innovation Solution
The method involves sharing pre-charged word lines and string selecting line groups/ground selecting lines to amortize time and energy consumption by reusing these pre-charged lines during the inference operation, utilizing a bit line controller and a word line and string selecting line controller to manage the switching of filters without switching individual word lines or string selecting line groups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If word lines and string selecting line groups are switched frequently during inference operation, then different filters can be selected, but time and energy consumption increase significantly
Solution Approach 1:
The patent pre-charges the word line and string selecting line groups before the inference operation begins. By performing this charging action in advance, the system eliminates the need for repeated charging cycles during filter switching, thereby reducing energy consumption while maintaining the ability to select different filters through line switching alone
Solution Approach 2:
The pre-charged word line and string selecting line groups serve multiple functions: they enable selection of different filters during inference operation without requiring re-charging, and they maintain their charged state to support multiple sequential operations. This multi-functionality reduces the overall energy footprint of the inference process
2Adaptability or versatility
If word lines and string selecting line groups are switched frequently during inference operation, then different filters can be selected, but operation time increases
Solution Approach 1:
The word line and string selecting line groups are pre-charged before the inference operation starts. This preliminary charging action eliminates the time penalty that would otherwise be incurred by charging these lines during each filter switch, allowing rapid filter selection without time loss
Solution Approach 2:
By maintaining the pre-charged state of the word line and string selecting line groups throughout the inference operation, the system enables continuous filter switching without interruption for re-charging. This continuity preserves the speed of operation while allowing versatile filter selection
3Measurement precision
If large amounts of data are processed during MAC operations, then accurate inference results are obtained, but energy consumption increases
Solution Approach 1:
The patent extracts the energy-intensive charging operation from the repeated filter switching cycle. By separating the one-time pre-charging action from the ongoing inference operations, the system maintains accurate processing of large data amounts during MAC operations while eliminating redundant energy consumption from frequent re-charging
Data Source
AI summary
An inference operation method and a controlling circuit of a 3D NAND artificial intelligence accelerator are provided. The 3D NAND artificial intelligence accelerator includes a plurality of memory cells, a plurality of bit lines, a plurality of word lines and a plurality of string selecting line groups each of which includes at least one string selecting line. The inference operation method includes the following steps: The patterns are inputted to the bit lines. The word lines are switched to switch the filters. The string selecting line groups are switched to switch the filters. In a word line pioneering scheme and a string selecting line group pioneering scheme, when the patterns inputted to each of the bit lines are switched, any one of the word lines is not switched and any one of the string selecting line groups is not switched.