Memristive Crossbar Arrays for Hyper-Dimensional In-Memory Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current hyper-dimensional computing systems face limitations in energy efficiency and scalability due to the need for extensive data movement between memory and processing units, and are restricted by the small data path and limited reprogrammability of existing implementations using memristive devices.
Innovation Solution
A system and method for hyper-dimensional computing that utilizes memristive devices in crossbar arrays for in-memory computing, allowing direct computation of hyper-dimensional vectors within memory units, eliminating the need for data movement and enabling efficient classification tasks with high dimensionality, and allowing for reprogrammability across various applications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If data is moved between memory systems and processing units in classical von-Neumann architecture, then computation can be performed, but energy consumption increases significantly
Solution Approach 1:
The patent merges memory and processing functions by implementing hyper-dimensional computing operations directly within the memory array. The crossbar array serves both as storage medium and computing engine, eliminating the need for data movement between separate memory and processing units. This integration resolves the fundamental energy efficiency problem of von-Neumann architecture while maintaining full computing capability.
Solution Approach 2:
The patent introduces memristive devices as intermediary elements that enable in-memory computing. These devices serve as the mediator between stored hyper-dimensional vectors and computation operations, allowing mathematical operations to be performed directly on the stored data without extraction to external processing units, thus resolving the energy-productivity contradiction.
2Productivity
If existing memristive device implementations are used, then in-memory computing is enabled, but the data path is limited to small dimensions (e.g., 32-bit)
Solution Approach 1:
The patent transitions from traditional small-dimensional (32-bit) data paths to hyper-dimensional spaces with dimensions in the thousands. By organizing memristive devices into large-scale crossbar arrays and using parallel column operations, the system achieves thousands of dimensions simultaneously, fundamentally expanding the data path width beyond conventional limitations without proportionally increasing complexity.
Solution Approach 2:
The patent segments the hyper-dimensional computing task into parallel column operations across the crossbar array. Each column processes a portion of the hyper-dimensional vectors simultaneously, allowing the system to handle thousands of dimensions through parallel segmentation rather than sequential processing, thus increasing effective data path width while managing complexity.
3Adaptability or versatility
If fixed implementations with limited ReRAM cells are used, then hardware simplicity is maintained, but reprogrammability and application versatility are restricted
Solution Approach 1:
The patent implements a universal hyper-dimensional computing platform where the same crossbar array and computational operations can be applied to multiple different applications and tasks. By using application-agnostic hyper-dimensional vector operations, the system achieves reprogrammability and versatility without requiring additional hardware resources, allowing a fixed number of ReRAM cells to serve multiple functions.
Solution Approach 2:
The patent introduces dynamic reconfigurability through software-controlled hyper-dimensional operations. The system can dynamically adapt to different applications by loading different hyper-dimensional vectors and operations into the fixed hardware architecture, enabling reprogrammability without physical reconfiguration or additional cells, thus achieving versatility with fixed hardware resources.
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 results in significant energy and area savings, enabling fast and efficient inference tasks with high accuracy, and supports a wide range of applications without degrading output quality, as demonstrated by experiments with 10,000-dimensional vectors.
Implementation Method 1
The item memory and the associative memory may be adapted for in-memory computing using memristive devices
Implementation Method 2
a system for hyper-dimensional computing for inference tasks. The device may comprise an item memory (IM) for storing hyper-dimensional item vectors and a query transformation unit connected to the item memory
Data Source
AI summary
A system for hyper-dimensional computing for inference tasks may be provided. The device comprises an item memory for storing hyper-dimensional item vectors, a query transformation unit connected to the item memory, the query transformation unit being adapted for forming a hyper-dimensional query vector from a query input and hyper-dimensional base vectors stored in the item memory, and an associative memory adapted for storing a plurality of hyper-dimensional profile vectors and for determining a distance between the hyper-dimensional query vector and the plurality of hyper-dimensional profile vectors, wherein the item memory and the associative memory are adapted for in-memory computing using memristive devices.