Butterfly Multiplexer for Homogeneous Sparse Bit Streams
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural inference accelerator hardware faces inefficiencies in compressing and decompressing multichannel bit streams due to serial processes that handle only one data bit stream at a time, leading to suboptimal power usage and storage efficiency in DRAM and SRAM.
Innovation Solution
A data-sparsity homogenizer system utilizing multiplexers and controllers to process multiple bit streams in parallel, rearranging data to make it more homogeneous, and a butterfly shuffler that packs and unpacks bit streams efficiently, allowing for random access and uniform sparsity across channels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If serial compression and decompression algorithms are used to handle one data bit stream at a time, then data compression is achieved, but processing speed and productivity are reduced
Solution Approach 1:
The patent divides the multichannel bit stream into multiple parallel channels, with each channel processed independently by dedicated compression and decompression circuits. This segmentation enables simultaneous processing of multiple data streams, dramatically increasing throughput while maintaining compression efficiency.
Solution Approach 2:
The patent combines multiple serial compression and decompression operations into a single parallel processing architecture. By merging the processing of multiple channels into concurrent operations, the system achieves both energy efficiency and high productivity.
2Quantity of substance
If data is compressed to reduce storage size in DRAM and SRAM, then storage requirements are reduced, but data access time and complexity increase
Solution Approach 1:
The patent performs preliminary organization of compressed data into structured formats with metadata that enables direct addressing. By pre-organizing the compressed bit streams with channel identifiers and position information, the system allows rapid random access without requiring sequential decoding.
Solution Approach 2:
The patent introduces intermediate buffer structures and control logic that mediate between the compressed storage and the processing units. These intermediaries enable efficient data retrieval by managing the conversion between compressed and uncompressed formats on-demand.
3Ease of operation
If non-homogeneous sparse data with clumped non-zero values is processed, then data representation is maintained, but processing efficiency and uniformity across channels are reduced
Solution Approach 1:
The patent applies different processing strategies to different regions of the data stream based on local characteristics. By identifying clumped non-zero values and applying targeted compression techniques to these regions while maintaining uniform data structures elsewhere, the system achieves both efficient representation and consistent processing across all channels.
Data Source
AI summary
A data-sparsity homogenizer includes a plurality of multiplexers and a controller. The plurality of multiplexers receives 2N bit streams of non-homogenous sparse data in which the non-homogenous sparse data includes non-zero value data clumped together. The plurality of multiplexers is arranged in 2N rows and N columns. Each input of a multiplexer in a first column receives a respective bit stream of the 2N bit streams of non-homogenized sparse data, and the multiplexers in a last column output 2N bit streams of sparse data that is more homogenous than the non-homogenous sparse data of the 2N bit streams. The controller controls the plurality of multiplexers so that the multiplexers in the last column output the 2N channels of bit streams of sparse data that is more homogeneous than the non-homogenous sparse data of the 2N bit streams.