Spatially Sparse Neural Network Accelerator for 3D Visual Analytics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network accelerators are inefficient in processing 3D sparse data due to their optimization for densely populated grids, leading to underutilization of hardware resources and high computational and memory requirements in 3D visual analytics applications.
Innovation Solution
The Spatially Sparse Neural Network Accelerator (SSpNNA) architecture, which includes a Weight plane based Active Voxel Execution Scheduler (WAVES) and Systolic and Multicast based MAC Computation (SyMAC), efficiently processes 3D sparse data by dynamically allocating resources and rearranging voxel feature maps based on weight planes, increasing compute resource utilization to approximately 90%.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional neural network accelerators optimized for densely populated grids are used to process 3D sparse data, then the hardware resources are underutilized and efficiency is low, but changing the architecture to handle sparse data efficiently increases device complexity
Solution Approach 1:
The accelerator architecture is segmented into specialized components: a weight plane-based active voxel execution scheduler that divides processing into weight planes, and systolic arrays that handle specific computation patterns. This segmentation allows efficient processing of sparse data by activating only relevant compute units for each weight plane, avoiding the need to process entire dense grids and thus improving productivity without requiring complete architectural redesign.
Solution Approach 2:
The execution scheduler dynamically allocates and activates compute resources based on the sparsity patterns and weight plane configurations. Rather than statically configuring all resources for dense processing, the system dynamically enables only the necessary processing elements for each operation, adapting to the actual data characteristics and achieving high efficiency for sparse workloads while maintaining compatibility with dense processing when needed.
2Quantity of substance
If conventional accelerators process 3D visual analytics data, then compute and memory requirements grow in cubic complexity with voxel resolution, but reducing resolution decreases measurement precision
Solution Approach 1:
The system extracts and processes only the non-zero or active voxels from the 3D data space, rather than processing the entire volumetric grid. By identifying and isolating the relevant active voxels and their corresponding weight planes, the accelerator eliminates computation on empty or irrelevant regions, reducing compute and memory requirements from cubic complexity to a function of the actual data sparsity while maintaining full voxel resolution for the active regions.
Solution Approach 2:
The accelerator applies different processing qualities to different regions of the 3D space: active voxels receive full-precision processing with dedicated computational resources, while inactive or zero-valued regions are skipped entirely. This local differentiation allows the system to maintain high measurement precision where needed (at active voxel locations) while dramatically reducing overall computational burden by avoiding processing in inactive regions.
Data Source
AI summary
Systems, apparatuses and methods may provide for technology that decodes data via an instruction that indicates a number of rulebooks to be processed, an input feature size, an output feature size, and a plurality of feature map base addresses, rearranges spatially distributed voxel output feature maps in the decoded data based on weight planes, and performs a channel-wise multiply-accumulate (MAC) operation on the rearranged spatially distributed voxel output feature maps to obtain an output, wherein the channel-wise MAC operation is performed as partial accumulations by a plurality of processing elements.