Resource Resettable DNN Accelerator with Virtual Tiling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Early deep neural network accelerators are limited in their ability to adjust dataflow according to application requirements, leading to suboptimal performance and high power consumption, especially in intermittent computing environments.
Innovation Solution
A resource resettable deep neural network accelerator system that includes a memory layer with a scratchpad memory and a register file memory, along with a virtual tiling layer that allows for dynamic adjustment of resource allocation and power management by reconfiguring resources such as cores and memories during inference operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If early deep neural network accelerators are designed to perform only a limited dataflow, then device complexity is reduced, but adaptability deteriorates
Solution Approach 1:
The patent implements dynamic reconfiguration of the scratchpad memory layer, allowing the tiling size to be changed during runtime through a virtual tiling layer. This enables the accelerator to adapt to different dataflow requirements without requiring multiple fixed-configuration hardware designs, thus improving adaptability while maintaining reasonable device complexity.
2Power
If cloud computing is used for deep neural network, then calculation performance is improved, but power consumption deteriorates
Solution Approach 1:
The patent changes the operational parameters of the accelerator by introducing a virtual tiling layer that can dynamically adjust tiling sizes. This allows the system to optimize power consumption by selecting appropriate tiling configurations for different inference workloads, enabling efficient on-device processing without requiring cloud computing resources.
3Reliability
If resources are allocated for deep neural network inference, then inference accuracy is maintained, but power consumption increases
Solution Approach 1:
The patent implements partial resource allocation by allowing the virtual tiling layer to activate only the necessary portion of the scratchpad memory layer based on the current inference workload. This enables the system to maintain inference accuracy by allocating sufficient resources when needed while reducing power consumption by deactivating unused resources during low-workload periods.
Data Source
AI summary
A resource resettable deep neural network accelerator according to an embodiment of the present disclosure includes: a memory layer including a scratchpad memory layer configured to divide deep neural network parameter data (hereinafter, data) in an external memory layer into a plurality of tiles and to load the divided tiles, and a register file memory layer configured to load tiled data of the scratchpad memory layer; and a plurality of cores configured to process an inference operation for the data loaded in the register file memory layer, wherein the memory layer includes a virtual tiling layer added to a certain location for loading the tiled data from a previous memory layer so as to correspond to a specific tiling size.