Feature
· Lossless compression of 16-bit neural-network feature data, with a demonstrated compression ratio of 2.2:1.
·
·
Lossy compression of FP32 model filters at a 4:1 ratio, with a reported inference-accuracy difference of 0.2%.
·
·
Three-stage hardware filter-compression pipeline comprising quantization, representative-symbol assignment, and Huffman coding.
·
·
Min-max uniform symmetric quantization converting FP32 filter values into 16-bit integers.
·
·
Range-based representative-symbol assignment to increase redundancy before entropy coding.
·
·
Huffman coding to reduce the size of the resulting symbol stream.
·
·
Reported hardware compression latency of three clock cycles.
·
·
Project evaluation using 256-byte blocks corresponding to CXL 3.0 flit granularity.
Application
· Feature-map and filter storage in AI accelerator memory subsystems.
·
·
Compression of neural-network d
Business Area
AI processors and neural-network accelerators. CXL-based memory expansion and data-management solutions.
Category
Interface Controller & PHY > CXL
Interface Controller & PHY
Arithmetic & Mathematic IP
Tech Specs
-
IP Name :
CMSS (CXL Memory Sub-System) Compression IP
-
Provider :
Scalable Architecture Lab, Sungkyunkwan University (SKKU)
-
Foundry :
N/A
-
Technology :
FPGA
Deliverables
· Three-stage hardware filter-compressor implementation comprising quantization, representative-symbol assignment, and Huffman coding.
·
·
Quantization stage converting FP32 filter values into 16-bit integers using min-max uniform symmetric quantization.
·
·
C/C++ reference models for feature compression-size estimation and filter compression analysis.
·
·
Fixed and adaptive mantissa-masking software for evaluating lossy filter representations.
·
·
Python scripts for ResNet-50 feature extraction, filter modification, model-weight restoration, and inference evaluation.
·
·
Build and execution scripts for compression experiments, compression-ratio summaries, and inference-accuracy comparisons.
Validation Status
· Hardware compression architecture implemented as a three-stage pipeline with a reported latency of three clock cycles.
·
·
Compression-ratio and inference-accuracy evaluation performed using C/Python simulation and ResNet-50 data.
·
·
Lossless compression of 16-bit feature maps achieved a reported compression ratio of 2.2:1.
·
·
Lossy compression of FP32 filters achieved a 4:1 ratio with a reported inference-accuracy difference of 0.2%.
·
·
Modified filters were restored and used for inference to evaluate accuracy relative to the original model.
Availability
Available as a digital compression IP for neural-network data in CXL memory subsystems.
Benefits
· Reduces the storage space required for neural-network feature maps and model filters.
·
·
Decreases data transfer volume, helping relieve memory-bandwidth pressure in AI accelerator systems.
·
·
Uses block-based processing to support bounded compression units and block-level memory access.
·
·
Combines numerical precision reduction, representative-symbol assignment, and entropy coding to improve filter compression.
·
·
Provides lossless feature compression while evaluating lossy filter compression against inference-accuracy requirements.
·
·
Uses a three-stage hardware pipeline with a reported compression latency of three clock cycles.
·
·
Provides software evaluation tools for examining the relationship between compression ratio, numerical precision, and inference accuracy.