Reconfigurable Integrated Circuit Using Adaptive Neural Network Cell
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing integrated circuit design methods require new metal routing mask layers to be produced when changes are identified late in the design process, and they are limited by the restricted number of spare cells available, which restricts the number of corrections that can be made in response to an Engineering Change Order (ECO).
Innovation Solution
A reconfigurable integrated circuit comprising a plurality of active cells and an adaptive neural network cell, which is coupled to one or more active cells and configured to implement a neural network with specific weights and biases, allowing for target circuit modifications without requiring changes in the physical masks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional spare cells are used for ECO corrections, then the circuit can be modified late in the design process, but the number of corrections is limited by the restricted number of spare cells available
Solution Approach 1:
The neural network cell is designed to be universally applicable for multiple types of logic corrections. By configuring the same physical cell with different weights and biases, it can implement various logic functions (AND, OR, NAND, NOR, XOR, etc.), thereby increasing the number of possible corrections without adding more spare cells to the circuit.
Solution Approach 2:
The invention changes the parameters (weights and biases) of the neural network cell to achieve different logic functions. By adjusting these parameters, the same cell can be reconfigured to correct different types of logic errors, effectively increasing the correction capacity without increasing the number of physical cells.
2Adaptability or versatility
If new metal routing mask layers are produced for corrections, then the circuit functionality can be changed, but the production time and cost increase significantly
Solution Approach 1:
The invention makes the circuit dynamically reconfigurable through software-based neural network configuration rather than static physical mask changes. The neural network cell can be reconfigured by loading different weight and bias values, allowing rapid adaptation without the time-consuming process of producing new metal mask layers.
Solution Approach 2:
The invention replaces the mechanical/physical process of producing new metal mask layers with a digital/software-based configuration process. Instead of physically altering the circuit through new mask layers, the correction is achieved by programming different parameters into the neural network cell, significantly reducing production time.
3Ease of repair
If traditional spare cells are used, then ECO changes can be implemented, but new metal routing mask layers are required which increases cost
Solution Approach 1:
The invention replaces the expensive mechanical process of producing new metal mask layers with a low-cost digital configuration process. The same physical cell can be reconfigured multiple times by loading different parameter sets, eliminating the need for costly remanufacturing of mask layers for each ECO.
Solution Approach 2:
The invention allows the same neural network cell to be reused for multiple different corrections by discarding one configuration and recovering/reusing the cell with a new configuration. This maximizes the utilization of existing hardware resources and avoids the cost of creating new physical structures for each correction.
Data Source
AI summary
A reconfigurable integrated circuit is presented. The reconfigurable integrated circuit (300) includes a plurality of active cells (302); and an adaptive neural network cell (304) coupled to one or more active cells. The adaptive neural network cell (304) has a digital neural network circuit (306) implementing a first neural network (308) having a plurality of weights and biases. For specific weights and biases and specific neural network inputs, the first neural network (308) provides specific neural network outputs. The adaptive neural network cell (304) is configurable with a set of target weights to achieve a target circuit modification.