Spintronic Approximate Memory for Bitwise Neural Weight Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI hardware accelerators face challenges in efficiently storing neural network weights with limited resources, leading to high power consumption and accuracy loss due to methods like low supply voltage SRAM and low refresh rate DRAM, which lack bit-level control and are affected by fabrication variations.
Innovation Solution
A spintronic adaptive approximate memory (SAAM) utilizing the stochastic behavior of magnetic tunnel junctions (MTJ) with adaptive approximation and bitwise control, combined with an approximation-aware learning algorithm, to reduce power consumption and area overhead while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If low supply voltage SRAM or low refresh rate DRAM is used for approximate memory, then power consumption is reduced, but area overhead increases and bit-level control is lost
Solution Approach 1:
The patent applies local quality by enabling bit-level control of approximation for different memory locations. Each bit position can be independently configured with different approximation levels, allowing critical bits to maintain high precision while less critical bits use approximate storage, thus optimizing the trade-off between area and accuracy without sacrificing overall system functionality
2Use of energy by stationary object
If low supply voltage SRAM or low refresh rate DRAM is used for approximate memory, then power consumption is reduced, but fabrication variations affect performance
Solution Approach 1:
The patent implements dynamics through adaptive approximation where the approximation level is dynamically adjusted based on input data characteristics and bit position. The system can switch between different approximation modes (e.g., bit-level approximation, word-level approximation) depending on the operational context, allowing the memory to maintain reliable performance across fabrication variations while optimizing power consumption
3Use of energy by stationary object
If quantization is applied to reduce weight storage bits, then area and power consumption are reduced, but network accuracy drops significantly
Solution Approach 1:
The patent applies parameter changes by introducing a controllable approximation parameter that can be adjusted based on bit position and data characteristics. Instead of uniform quantization, the system varies the approximation level as a parameter across different bit positions, allowing finer precision for critical bits and coarser approximation for less critical bits, thus maintaining network accuracy while reducing overall power consumption
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
The SAAM achieves up to 40% power efficiency improvement with negligible accuracy loss, offering improved error rates and reduced area overhead compared to conventional methods.
Implementation Method 1
Magnetic tunnel junction (MTJ), as the basic element of the spintronic circuit, also has a very special property that can be useful in approximate computing and approximate memory implementation [12]. This feature is the stochastic behavior of MTJ in the current below the critical current [13].
Data Source
AI summary
A spintronic adaptive approximate memory (SAAM) is used with or in a neural network. A methodology adaptive approximation is used based on the stochastic behavior of magnetic tunnel junctions (MTJ). The SAAM offers an innovative solution for storing neural network weights with input data-dependent controlled accuracy. The memory significantly reduces power consumption and area overhead while minimizing the loss of accuracy. Key features include bitwise control over memory accuracy, an approximation-aware learning algorithm, and an adaptive write circuit which broadens applications of the designed approximate memory, beyond neural network hardware accelerators. The disclosure evaluates SAAM's performance through functional simulations, statistical analyses, and neural network implementations, demonstrating its advantages over existing approximate memories. These simulations demonstrate power efficiency improvements ranging from 10% to 40%. This enhancement is achieved at the cost of a negligible accuracy reduction, ranging from 1% to 7%.