Neural Network Unit Binary Search Read Method
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network units require exponentially increasing steps for N-bit sensing, leading to a significant increase in read time as resolution increases.
Innovation Solution
The proposed binary search read method uses multiple voltage-controlled current source and single-bit sense amplifier circuits to sense the BL current level with N-bit resolution through N-step binary search, resulting in a linear increase in read time with resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional N-bit sensing method is used, then measurement precision is improved, but read time exponentially increases
Solution Approach 1:
The sensing process is segmented into N sequential steps, each step determining one bit of the N-bit resolution through binary search. Instead of requiring 2^N-1 steps for N-bit sensing, the patent divides the sensing task into N manageable steps where each step narrows down the possible current levels by comparing against a reference level, achieving linear complexity O(N) rather than exponential complexity O(2^N).
Solution Approach 2:
The sensing circuit uses feedback from single-bit sense amplifier comparisons to guide the binary search process. The sense amplifier compares the bit line current to a reference current, and the result feeds back to determine which half-range to search next, enabling efficient N-step binary search that achieves N-bit resolution with linear time complexity.
2Measurement precision
If N-bit resolution sensing is implemented, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The sensing circuit is designed with multi-functionality to handle N-bit resolution sensing through a universal binary search approach. The same reference synapse and sense amplifier circuit are reused across N steps, with the reference current being adjusted at each step rather than requiring N separate reference circuits. This universal design achieves high resolution without proportionally increasing device complexity.
Solution Approach 2:
The patent achieves N-bit resolution by dynamically changing the reference current parameter at each step of the binary search. Instead of using fixed reference circuits, the reference synapse current is adjusted based on the search progress, allowing the same physical circuit to perform N-bit sensing through parameter modulation rather than requiring N times more circuit elements.
3Loss of time
If binary search read method is used, then read time complexity is reduced, but manufacturing precision requirements increase
Solution Approach 1:
The binary search method uses partial action by only requiring N steps instead of the full 2^N-1 steps needed for exhaustive search. This partial approach achieves sufficient precision for neural network operations while significantly reducing the manufacturing precision requirements, as each step only needs to distinguish between two ranges rather than precisely calibrate all 2^N possible levels.
Data Source
AI summary
A neural network unit is disclosed. This neural network unit includes: a main synapse array having a plurality of main synapses in rows and columns of the array, each connected to one or more bit lines associated with the network unit; a main synapse driver for applying an input to one or more of the main synapses in a row; a reference synapse for generating a particular output to a selected one or more associated bit lines; and, a sensing circuit for controlling the reference synapse in generating the output in a selected column of the array, wherein the sensing circuit is adapted to infer an output from one or more main synapses in a selected column by: defining a reference range of the output to be generated by reference synapse; setting a reference level from the reference range; activating a selected reference synapse to implement the reference level that is set or adjusted; and comparing an output from the main synapse summed with the output from the activated reference synapse in the selected column with a threshold value predetermined.