Multi-Dimensional PE Interconnect for Faster Data Movement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional calculation devices with multiple processing elements (PEs) arranged in a mesh structure face inefficiencies in data movement, leading to prolonged processing times for operations like Sum and Max processing, which hinders speed improvements and increases power consumption.
Innovation Solution
The calculation device employs a coordinate system with X, Y, and additional dimensions (Z and W) to enable data input/output between processors that are not adjacent in the traditional X and Y directions, allowing for data movement through diagonal and non-adjacent connections, thereby reducing the number of required processors and optimizing wiring length.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If data is moved between adjacent PEs in a mesh structure using conventional methods, then the system maintains simple wiring and easy operation, but the data movement speed is slow and processing time is prolonged
Solution Approach 1:
The patent extends the conventional two-dimensional mesh structure to a four-dimensional coordinate system (X, Y, Z, W dimensions). This allows processors to communicate not only with adjacent processors in the traditional X and Y directions but also with processors in diagonal and non-adjacent positions through the additional Z and W dimensions, significantly reducing the number of data movement steps required for aggregation operations
2Productivity
If multiple data movement operations are performed in sequential order in conventional mesh structures, then the system maintains simple processor arrangement, but the total processing time increases and power consumption rises
Solution Approach 1:
By introducing Z and W dimensions to the coordinate system, the patent enables processors to reach distant processors in fewer steps. For example, in Sum processing operations, the number of data movement steps is reduced from 16 times in conventional mesh structures to just 4 times in the four-dimensional structure, directly reducing processing time and improving productivity
Solution Approach 2:
The patent pre-organizes processors in a four-dimensional coordinate system before data movement operations begin. This preliminary arrangement allows processors to directly access non-adjacent processors through the extended dimensions, eliminating the need for multiple sequential data movement operations and reducing overall processing time
3Use of energy by moving object
If data is moved through multiple sequential operations in conventional accelerators, then the system maintains simple processor interconnection, but power consumption increases and energy efficiency decreases
Solution Approach 1:
The four-dimensional coordinate system reduces the number of data movement operations required for aggregation processing from 16 steps to 4 steps in conventional mesh structures. Since each data movement operation consumes power, reducing the number of steps by 75% directly reduces power consumption while simultaneously improving calculation speed
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An accelerator (10) includes multiple PEs (12). The multiple PEs (12) are represented by a coordinate system, which includes two dimensions of X direction and Y direction and two or more different dimensions indicating different directions. Each PE (12) is capable of performing data input or data output with another PE (12) adjacent in the X direction or the Y direction, and is also capable of performing data input or data output with another PE (12) adjacent in different dimension.