Research map / Architectures and efficiency / RNNs and normalization
"L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training"
Jonghyun Bae, Woohyeon Baek, Tae Jun Ham, Jae W. Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add patch-level parallelism to the L3 decoder (ablation) | L3 decoder, Cityscapes · HD/FHD/UHD | decode time: reduces the decoding time by 76.8%, 83.9%, and 86.5% for HD, FHD, and UHD | Improved |
| add pixel-wise parallel base-delta decoding (ablation) | L3 decoder, Cityscapes · HD/FHD/UHD | decode time: reduces the decoding time by an average of 10.8% | Improved |
| use the custom Paeth filter (top-left, top, top-right) with row-wise parallel decode (ablation) | L3 decoder, Cityscapes · HD/FHD/UHD | decode time: reduces the decoding time by an average of 46.0% | Improved |
| combine pixel-wise base-delta decoding and custom Paeth filter (ablation) | L3 decoder, Cityscapes · HD/FHD/UHD | decode time: 49.5%, 56.7%, and 59.1% from the Baseline for HD, FHD, and UHD | Improved |
| replace PNG with L3 for lossless data preparation | DNN training pipeline · Cityscapes HD/FHD/UHD | data preparation throughput: 5.67x, 9.29x, and 15.71x for HD, FHD, and UHD | Improved |
| replace lossless WebP with L3 | DNN training pipeline · Cityscapes HD/FHD/UHD | data preparation throughput: 2.41x, 3.25x, and 4.51x | Improved |
| replace PNG with L3 | end-to-end training of seven models · Cityscapes and KITTI | training throughput: 1.71x geomean speedup | Improved |
| replace lossy JPEG with L3 at equal accuracy | end-to-end training of seven models · Cityscapes and KITTI | training throughput: 1.25x geomean, up to 1.77x | Improved |
| replace lossy WebP with L3 at equal accuracy | end-to-end training of seven models · Cityscapes and KITTI | training throughput: 1.93x geomean, up to 2.87x | Improved |
| use default-quality JPEG (Q75) instead of lossless PNG | DNN training on PNG vs JPEG · DDRNet23-slim, Cityscapes | mIoU: 0.729 -> 0.693 | Hurt |
| use default-quality JPEG (Q75) instead of lossless PNG | DNN training on PNG vs JPEG · MaskFormer, Cityscapes | mIoU: 0.685 -> 0.669 | Hurt |
| use JPEG (Q75) instead of PNG | DeepLabv3+ · Cityscapes | mIoU: 0.801 -> 0.801 | No clear effect |
Closest papers by meaning
- DepthShrinker: A New Compression Paradigm Towards Boosting Real-Hardware Efficiency of Compact Neural Networks (ICML 2022 · similarity 0.85)
- FireCaffe: Near-Linear Acceleration of Deep Neural Network Training on Compute Clusters (CVPR 2016 · similarity 0.85)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · similarity 0.84)
- 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11) (NeurIPS 2025 · similarity 0.84)
- Poplar: Efficient Scaling of Distributed DNN Training on Heterogeneous GPU Clusters (AAAI 2025 · similarity 0.83)
- DeepSpeed Data Efficiency: Improving Deep Learning Model Quality and Training Efficiency via Efficient Data Sampling and Routing (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
