Processor Memory Access for Random Reads on Compressed Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing systems face limitations in memory bandwidth, known as the 'memory-wall,' which restricts data transfer between processor and memory, and existing data compression methods require decompression of entire structures, increasing memory bandwidth and processor load.
Innovation Solution
A processor architecture that allows direct operation on compressed data without decompression, using a value map and integrated compression/decompression logic within memory access structures, enabling random-access to individual values within compressed structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If data compression is used to reduce memory bandwidth requirements, then memory bandwidth consumption is reduced, but data must be decompressed before processing which requires additional memory bandwidth and processor resources
Solution Approach 1:
The patent combines the decompression logic directly within the memory access structures (load/store units), merging the compression/decompression function with the memory access path. This eliminates the need for separate decompression stages and allows compressed data to be processed directly from memory without requiring additional buffer memory or separate decompression processing cycles.
Solution Approach 2:
The patent introduces an intermediary value map data structure that stores metadata about the compressed data structure (such as non-zero element positions in sparse matrices). This value map enables the processor to directly access and decompress only the specific elements needed for computation, rather than requiring full decompression, thereby reducing the decompression overhead while maintaining the memory bandwidth benefits of compression.
2Ease of operation
If entire compressed structures are decompressed for processing, then data can be operated upon, but additional memory capacity and processor register file pressure increase
Solution Approach 1:
The patent extracts only the necessary elements from the compressed data structure for processing by using the value map to identify and retrieve specific elements (e.g., non-zero elements in sparse matrices) directly from memory into processor registers. This selective extraction avoids the need to load and decompress entire compressed structures or maintain large uncompressed buffers, thereby reducing memory capacity requirements while maintaining full data processing capability.
3Quantity of substance
If traditional compressed memory systems are used, then data storage efficiency improves, but the memory wall limitation persists due to decompression bottlenecks
Solution Approach 1:
The patent performs preliminary actions by pre-computing and storing the value map (metadata about the compressed structure) alongside the compressed data in memory. This preliminary organization of data with its structural information enables the processor to directly access and decompress only the necessary elements during computation, eliminating the decompression bottleneck and maximizing effective memory bandwidth while maintaining high data storage efficiency.
Data Source
AI summary
The present application provides a method of randomly accessing a compressed structure in memory without the need for retrieving and decompressing the entire compressed structure.