CNN Compression Circuit Using Bit Maps to Cut DRAM Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional Neural Networks (CNNs) face performance slowdowns due to high DRAM bandwidth consumption from exchanging high-dimensional intermediate vectors/tensors, which is not efficiently managed by existing data compression techniques.
Innovation Solution
A high throughput hardware unit is implemented, comprising a compression circuit that reduces data size using a bit map and non-zero values representation, achieving efficient lossless data compression, specifically designed for CNNs, which can be integrated into intermediate layers and reduce DRAM bandwidth usage without affecting detection results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If high dimensional intermediate vectors/tensors are exchanged through DRAM in CNNs, then computation capability is improved, but DRAM bandwidth consumption increases significantly
Solution Approach 1:
The patent segments the high-dimensional intermediate tensors into multiple coding blocks, where each block is independently compressed using bit maps. This segmentation allows the system to process and compress data in manageable units, reducing the overall DRAM bandwidth requirement while maintaining computation capability.
Solution Approach 2:
The patent extracts and removes redundant information from the intermediate tensors by using bit maps to represent only the necessary data elements. By taking out and eliminating duplicate or unnecessary data, the system reduces DRAM bandwidth consumption without losing essential computational information.
2Quantity of substance
If data compression is applied to reduce DRAM bandwidth, then bandwidth consumption is reduced, but compression complexity increases
Solution Approach 1:
The patent uses simple bit map structures as disposable compression objects that are easy to generate and discard. Each coding block is compressed using a straightforward bit map representation, avoiding complex compression algorithms while achieving significant bandwidth reduction. The simplicity of the bit map approach keeps compression complexity low.
3Quantity of substance
If compression ratio is increased to reduce data size, then storage efficiency is improved, but lossless compression becomes more difficult to maintain
Solution Approach 1:
The patent creates a simplified copy of the original data using bit maps that preserve the essential information needed for lossless reconstruction. The bit map serves as a compact representation that can be easily converted back to the original data, ensuring lossless compression while achieving high compression ratios.
Data Source
AI summary
An apparatus includes a first memory interface circuit, a second memory interface circuit, and a compression circuit coupled between the first memory interface circuit and the second memory interface circuit. The compression circuit may be configured to receive a coding block of data via the first memory interface circuit, generate a reduced size representation of the coding block, and write the reduced size representation of the coding block to an external memory using the second memory interface circuit. The reduced size representation of said coding block generally comprises a first bit map, a second bit map, and zero or more non-zero values.