Research map / Graph learning / Graph algorithms and clustering
AutoGraph: Optimizing DNN Computation Graph for Parallel GPU Kernel Execution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · Inception-v3 | latency: 2.928 -> 2.799 ms | Helped |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · ResNet-50 | latency: 1.988 -> 1.905 ms | Helped |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · ResNeXt-50 | latency: 3.056 -> 2.892 ms | Helped |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · NasNet-A | latency: 6.483 -> 5.850 ms | Helped |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · NasNet-Mobile | latency: 3.821 -> 1.883 ms | Helped |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · RNNTC-SRU | latency: 0.387 -> 0.387 ms | No clear effect |
| Optimize the computation graph for multi-stream GPU execution using mixed critical-path cost and measured runtime feedback | DNN inference graphs · BERT | latency: 6.473 -> 6.240 ms | Helped |
| Use dynamic programming to select an optimization partition rather than directly using the minimum partitioning set (ablation) | AutoGraph · benchmark DNN models | latency | Helped |
| Disable graph optimization and directly measure the initial computation graph (ablation) | AutoGraph · benchmark DNN models | latency | Hurt |
| Preserve parallel branches instead of applying all three substitutions in the NasNet cell (ablation) | AutoGraph · NasNet-A and NasNet-Mobile | latency: 13.086 ms -> 5.850 ms; 6.540 ms -> 1.883 ms | Helped |
| Use a mixture of group numbers 8 and 4 rather than group number 8 for all ResNeXt convolutions (ablation) | ResNeXt-50 · ResNeXt-50 | latency: 5.933 ms -> 2.892 ms | Helped |
Closest papers by meaning
- AutoGO: Automated Computation Graph Optimization for Neural Network Evolution (NeurIPS 2023 · similarity 0.84)
- Poplar: Efficient Scaling of Distributed DNN Training on Heterogeneous GPU Clusters (AAAI 2025 · similarity 0.84)
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.84)
- StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning (ICML 2026 · similarity 0.83)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.82)
- RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs (ICLR 2025 · similarity 0.82)
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.