Neural Network Memory Device Adaptive Program Pulse Time
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks face challenges with growing model size during training, particularly due to high costs, limited memory density, and leakage power issues with DRAM, while phase-change memory (PCM) offers advantages but suffers from low write performance, high energy consumption, and low endurance.
Innovation Solution
A memory device with a controller that determines computation duration in neural network layers to select appropriate program operations, such as rapid set or full set operations, to optimize program pulse time and data retention, adapting to different computation durations and bit importance across layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DRAM is used as main memory for neural networks, then write performance is high and energy consumption is low, but cost is high, scaling is difficult, memory density is limited, and leakage power is significant
Solution Approach 1:
The patent applies parameter changes by transitioning from DRAM to PCM technology, fundamentally changing the physical state and operational parameters of the memory system. PCM operates without leakage power and provides higher memory density while maintaining adequate write performance for neural network workloads, effectively resolving the contradiction between write performance and memory density through a different technological parameter regime
2Quantity of substance
If PCM is used as main memory for neural networks, then cost is low, scaling is easy, memory density is high, and leakage power is eliminated, but write performance is low, energy consumption is high, and endurance is low
Solution Approach 1:
The patent applies local quality by implementing layer-aware differentiation, where different program operations are selectively applied to different neural network layers based on their specific computation durations and data retention requirements. This localized approach optimizes write performance for each layer's specific needs while maintaining the overall benefits of PCM technology
Solution Approach 2:
The patent applies dynamics through adaptive program operation selection that dynamically adjusts the programming strategy based on real-time computation duration measurements. The system dynamically chooses between rapid set and full set operations, and determines refresh intervals based on actual data retention characteristics, thereby optimizing write performance adaptively rather than using a static approach
3Duration of action of stationary object
If longer program pulse time is used in PCM, then data retention is improved, but energy consumption increases and write performance decreases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the program pulse time parameter based on the specific computation duration requirements of different neural network layers. Rather than using a fixed long pulse time for all data, the system adapts the pulse duration parameter to match actual retention needs, thereby reducing unnecessary energy consumption while maintaining adequate data retention
Solution Approach 2:
The patent applies partial action by using rapid set operations (shorter program pulse time) for data that does not require long retention, and only using full set operations (longer program pulse time) when actually needed. This partial application of the longer pulse time approach reduces overall energy consumption while maintaining data retention for critical data points
4Duration of action of stationary object
If frequent refresh operations are performed on activation maps with long computation duration, then data retention is maintained, but energy consumption increases
Solution Approach 1:
The patent applies feedback by measuring actual data retention characteristics and using this information to determine appropriate refresh intervals. The system continuously monitors and adjusts refresh timing based on observed retention behavior, ensuring data is refreshed only when necessary rather than following a fixed frequent schedule, thereby reducing energy consumption while maintaining data integrity
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 enhances write performance, reduces energy consumption, and improves endurance while maintaining data precision, effectively addressing the limitations of existing memory technologies for neural networks.
Implementation Method 1
A memory device including: a memory array used for implementing neural networks (NN)... phase-change memory (PCM) is proposed as an alternative main memory device
Data Source
AI summary
A memory device includes: a memory array used for implementing neural networks (NN), the NN including a plurality of layers; and a controller coupled to the memory array, the controller being configured for: determining a computation duration of a first data of a first layer of the plurality of layers; selecting a first program operation if the computation duration of the first data of the first layer is shorter than a threshold; and selecting a second program operation if the computation duration of the first data of the first layer is longer than the threshold, wherein the second program operation has a longer program pulse time than the first program operation.