NAND In-Memory Computing with Current Accumulation for Higher SNR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In-memory computing for artificial neural networks is computationally intensive due to large data transfers and signal-to-noise issues in vector dot product computations, particularly in NAND-based systems where current differences are small and prone to noise.
Innovation Solution
Accumulating currents from multiple NAND strings to improve signal-to-noise ratio by programming threshold voltages of memory cells to represent vector values, allowing for in-memory compute operations that enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If current differences are used for vector dot product computations in NAND-based systems, then in-memory computing capability is achieved, but signal-to-noise ratio deteriorates due to small current differences being prone to noise
Solution Approach 1:
The patent combines multiple NAND strings into a single computational unit, where currents from multiple strings are accumulated and added together. This merging approach transforms individual small current differences into a larger combined current signal, improving the signal-to-noise ratio while maintaining in-memory computing capability. The accumulation of currents from multiple strings effectively amplifies the computational signal beyond what a single string could provide.
2Measurement precision
If weights are stored in memory and transferred to processing units for computation, then computational accuracy is maintained, but data transfer intensity and power consumption increase
Solution Approach 1:
The patent enables the memory system to perform computations autonomously without requiring data transfer to external processing units. By implementing in-memory computing where NAND strings directly compute vector dot products through current accumulation, the system serves its own computational needs, eliminating the energy-intensive data transfer between memory and processing units while maintaining computational accuracy.
3Productivity
If multiple NAND strings are used for computation, then computational throughput is improved, but device complexity increases
Solution Approach 1:
The patent designs NAND strings to serve multiple functions: they simultaneously act as data storage elements, computational units for vector dot products, and signal generation sources. This multi-functionality allows multiple NAND strings to be used for computation without proportionally increasing system complexity, as each string is already designed to handle multiple operational modes. The same hardware infrastructure supports both storage and computation tasks.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Improves accuracy and efficiency of in-memory computing by enhancing the signal-to-noise ratio in NAND-based systems, reducing computational intensity and power consumption.
Implementation Method 1
Accumulating currents from multiple NAND strings to improve signal-to-noise ratio
Data Source
AI summary
Technology for NAND in-memory computing. Currents from multiple NAND strings may be accumulated in order to improve the signal-to-noise ratio to thereby improve accuracy for in-memory compute using NAND. The memory system may program threshold voltages of memory cells of a number of NAND strings to represent a corresponding number of copies of a first set of values such as a first vector. Voltages may be applied to gates of the memory cells in order to represent a second set of values (e.g., a second vector). The current from each NAND string is accumulated at a sense node resulting in an “accumulated signal.” The foregoing may be applied for another set of NAND strings and another sense node to provide a second accumulated signal. The accumulated signals may be compared to determine a result for an in-memory compute (e.g., vector/vector multiply).