Tensor Transform Apparatus Local Processing Bandwidth Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for performing transform operations on data tensors require frequent data transmission between processing resources and host processors, leading to bandwidth and processing resource inefficiencies.
Innovation Solution
A data processing apparatus and method that enables transform operations to be performed directly on processing resources, utilizing communication circuitry, sub-block processing circuitry, and transformation circuitry to divide and process tensors in independent blocks, reducing the need for extensive data transfer and allowing parallel processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If transform operations are performed on the host processor, then the processing resource can obtain the transformed data tensor, but bandwidth and processing resources on the host processor are consumed
Solution Approach 1:
The data tensor is divided into multiple blocks, where each block can be independently processed. This segmentation allows the processing resource to perform transform operations on individual blocks locally without requiring the entire tensor to be transferred to the host processor, thereby reducing bandwidth consumption and host processor load while maintaining processing capability.
Solution Approach 2:
The processing resource is equipped with dedicated processing circuitry that can perform transform operations autonomously without requiring host processor intervention. This self-service capability allows the processing resource to handle its own data transformation needs, eliminating the need for data round-trips to the host processor and reducing both bandwidth usage and host processor resource consumption.
2Adaptability or versatility
If data is transmitted back and forth between processing resource and host processor, then transform operations can be executed, but bandwidth is consumed
Solution Approach 1:
The patent introduces a new dimension of processing by adding dedicated processing circuitry at the processing resource level, moving transform operations from the host processor dimension to the edge processing dimension. This dimensional shift enables local execution of transform operations, eliminating the need for extensive data transmission between the processing resource and host processor while preserving full transform operation capability.
3Productivity
If the entire data tensor is processed at once, then transform operations can be completed, but processing efficiency is reduced due to lack of parallelism
Solution Approach 1:
The data tensor is divided into multiple independent blocks that can be processed in parallel. Each block is handled by the processing circuitry independently, enabling simultaneous execution of transform operations across multiple blocks. This segmentation increases processing throughput by utilizing parallel processing capabilities while the modular block-based approach keeps the processing architecture manageable and scalable.
Data Source
AI summary
The present disclosure relates to a data processing apparatus for a processing resource to perform a transform operation on an input tensor for the processing resource, said input tensor being formed of a plurality of blocks, each block being a portion of said input tensor capable of being operated on independently of each other, said data processing apparatus comprising: communication circuitry to communicate with a control module and a shared storage of said processing resource; processing circuitry to perform said transform operation, said processing circuitry comprising sub-block processing circuitry and transformation circuitry; and a local storage to store transform operation output from said processing circuitry; wherein said communication circuitry is configured to: receive one or more transform parameters; read a first input sub-block from said shared storage, said first input sub-block being a portion of a first block of said input tensor corresponding to a processing unit of said processing circuitry; and write a first output sub-block to said shared storage, wherein said sub-block processing circuitry is configured to: divide said first block of said input tensor into one or more input sub-blocks capable of being operated on independently of each other based on said one or more transform parameters; and wherein said transformation circuitry is configured to: perform said transform operation on said first input sub-block based on said one or more transform parameters to generate said first output sub-block; and write said first output sub-block to said local storage.