Neural Network Processor With 3D Memory Shifting for Vector Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vector processing methods in autonomous vehicle systems are inefficient when applying depth-oriented vectors on two-dimensional input values, leading to suboptimal utilization of 3D memory units.
Innovation Solution
Employing a 3D memory element that allows for shifting data across rows, enhancing access to depth-oriented vector data and improving throughput by facilitating both 2D and 3D operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If depth-oriented vectors are applied on two-dimensional input values using existing vector processing methods, then processing can be performed, but the utilization of 3D memory units is suboptimal and processing efficiency is low
Solution Approach 1:
The patent transforms two-dimensional input data into three-dimensional representations by introducing a depth dimension. This allows 3D memory units to be fully utilized for storing and processing depth-oriented vectors, thereby improving both processing efficiency and memory utilization. The conversion enables the system to leverage the full capacity of 3D memory structures rather than leaving them underutilized.
Solution Approach 2:
The patent designs a processing system that can handle both 2D and 3D operations using the same hardware infrastructure. By creating a unified approach that works with both two-dimensional and three-dimensional data structures, the system achieves multi-functionality, allowing 3D memory units to serve multiple purposes including storing depth information, intermediate results, and final outputs.
2Quantity of substance
If 3D memory units are used for storing depth-oriented vector data, then data capacity is increased, but data access efficiency deteriorates without proper shifting mechanisms
Solution Approach 1:
The patent introduces dynamic shifting mechanisms that allow data to be moved between different regions of the 3D memory unit as needed. This dynamic reconfiguration enables the system to optimize data access patterns by positioning frequently accessed data in easily reachable locations while maintaining the large storage capacity of the 3D structure. The shifting capability makes the memory system adaptive to different access requirements.
Solution Approach 2:
The patent employs intermediate buffer structures and shifting mechanisms that act as mediators between the large-capacity 3D memory and the processing units. These intermediaries facilitate efficient data transfer by pre-positioning data in optimal locations before processing, thereby maintaining high data access speed despite the increased storage capacity.
3Productivity
If conventional vector processing is used, then implementation is simpler, but throughput is reduced
Solution Approach 1:
The patent divides the processing architecture into specialized segments including separate units for 2D-to-3D conversion, depth-oriented vector processing, and result aggregation. This segmentation allows each component to be optimized for its specific function, thereby achieving high throughput. While the overall system becomes more complex, each individual segment remains relatively simple and well-defined.
Solution Approach 2:
The patent performs preliminary transformations of input data into 3D representations before the main processing stage. By pre-converting 2D input data into 3D depth-oriented vectors and pre-positioning data in the 3D memory unit, the system prepares data in advance for efficient processing, thereby increasing throughput without requiring complex real-time transformations during the critical processing path.
Data Source
AI summary
A method for neural network processing, the method may include applying, by a neural network processor, a group of neural network kernels of respective layers of a neural network to provide neural network output results. The applying may include applying a (2D) kernel of the group of kernels on 2D layer input values of a first layer of the neural network to provide first layer output values. The applying may include scanning the 2D layer input values with the 2D kernel. The scanning may include flattening the 2D kernel, flattening the 2D layer input values, and storing first layer output values in a 3D memory unit. The applying may also include applying a 3D kernel of the group of kernels, on 3D layer input values of a second layer of the neural network to provide second layer output values, wherein the applying comprises scanning at least one feature vector of the 3D layer input values with at least one vector of the 3D kernel.