Adaptive Quantization for Analog In-Memory Computing Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing in-memory computing (IMC) systems face challenges in robustness to noise and variations in large crossbar arrays, particularly in analog IMC systems, which affects the accuracy and efficiency of calculations.
Innovation Solution
The proposed method employs adaptive quantization of neural network parameters using magnetic memory devices (MMDs), where parameters are quantized based on a conductance shift sensing process, generating a conductance shift lookup table, and iteratively tuning the parameters to minimize calculation errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If analog in-memory computing systems use large crossbar arrays for calculations, then computational throughput and efficiency are improved, but robustness to noise and variations deteriorates, affecting calculation accuracy
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the quantization bit-depth based on the calculated error metrics. When error exceeds the threshold, the system increases quantization precision (changes the parameter of quantization bits), thereby improving calculation accuracy without permanently increasing system complexity. This adaptive parameter adjustment resolves the contradiction by allowing the system to maintain high throughput while compensating for noise sensitivity through selective precision enhancement.
Solution Approach 2:
The patent implements a feedback mechanism where the system continuously monitors calculation errors in the analog IMC system and uses this information to adjust quantization parameters. The error calculation unit computes differences between expected and actual results, and this feedback drives the adaptive quantization process. This closed-loop feedback resolves the contradiction by enabling the system to maintain high throughput while automatically compensating for noise-induced accuracy degradation through real-time parameter adjustment.
2Measurement precision
If higher precision quantization is applied to neural network parameters, then calculation accuracy is improved, but hardware complexity and resource requirements increase
Solution Approach 1:
The patent applies dynamics by making the quantization bit-depth adaptive rather than static. The system dynamically adjusts the number of bits used for quantization based on real-time error measurements and calculated thresholds. This dynamic approach resolves the contradiction by allowing the system to use higher precision only when and where needed, rather than uniformly across all operations, thereby maintaining accuracy while minimizing hardware complexity and resource utilization.
Solution Approach 2:
The patent changes the quantization parameter (bit-depth) based on error thresholds and performance requirements. Instead of using fixed high precision throughout the system, the patent adjusts the precision parameter adaptively, increasing it only when error exceeds acceptable levels. This parameter change strategy resolves the contradiction by enabling high accuracy when necessary while avoiding the constant hardware overhead associated with fixed high-precision quantization.
3Loss of energy
If existing hardware is used without adaptation, then infrastructure costs are reduced, but performance and throughput specifications cannot be met
Solution Approach 1:
The patent applies self-service by enabling existing hardware to adapt and optimize its own performance through the adaptive quantization mechanism. The system uses the available hardware resources and automatically adjusts quantization parameters to meet performance specifications without requiring additional hardware infrastructure. This self-service approach resolves the contradiction by allowing existing hardware to achieve higher throughput and performance through intelligent parameter adaptation rather than through hardware upgrades.
Solution Approach 2:
The patent changes operational parameters (quantization bit-depth, thresholds) of existing hardware to extract maximum performance from the available infrastructure. By dynamically adjusting these parameters based on error metrics and performance requirements, the system enables existing hardware to meet throughput specifications that would otherwise require more powerful or additional hardware. This parameter adaptation resolves the contradiction by achieving high performance through software-controlled parameter optimization rather than hardware expansion.
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
This approach reduces the multiplication and accumulation (MAC) errors in analog IMC systems, enhances the accuracy of deep neuron network models under significant memory device variations and circuit noise, and allows for more efficient use of existing hardware.
Implementation Method 1
determining a conductance error of the MMD, based on a specified state of the MMD and based on a sensed conductance of the MMD
Data Source
AI summary
One or more systems, methods and/or machine-readable mediums are described herein for adaptively quantizing the parameters that are aimed to be deployed on in-memory computing systems based on magnetic memory devices. The method includes quantizing parameters based on a conductance shift sensing process, a sensed conductance shift value, a conductance shift lookup table, and a parameter to be quantized. The parameter can be quantized by using a value recorded in the lookup table, such as rounding to the nearest value. The lookup table can be generated by the conductance shift sensing process which can enabling recording of a sensed conductance of a magnetic memory device and an associated device state. The conductance shift sensing process can set the magnetic memory device (MMD) to different states and can measure the conductance shift of the MMD, caused by the state setting, using suitable equipment.


