Research map / Architectures and efficiency / CNN design
Deep Tensor Convolution on Multicores
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 |
|---|---|---|---|
| Use the proposed fast convolution implementation instead of TensorFlow (Eigen 3.3) | 2D convolution on CPU · 224×224 ImageNet image, 3 convolution layers, 1 core | throughput (MVox/s): 0.18 -> 0.89 | Helped |
| Use the proposed fast convolution implementation instead of Caffe (Intel MKL) | 2D convolution on CPU · 224×224 ImageNet image, 3 convolution layers, 1 core | throughput (MVox/s): 0.19 -> 0.89 | Helped |
| Use the proposed fast convolution implementation instead of TensorFlow (Eigen 3.3) | 2D convolution on CPU · 224×224 ImageNet image, 3 convolution layers, 18 cores | throughput (MVox/s): 1.77 -> 10.9 | Helped |
| Use the proposed fast convolution implementation instead of Caffe (Intel MKL) | 2D convolution on CPU · 224×224 ImageNet image, 3 convolution layers, 18 cores | throughput (MVox/s): 0.41 -> 10.9 | Helped |
| Use AVX/FMA vectorized operations | Fast convolution implementation · Single-core convolution | speed-up: 32-fold | Helped |
| Parallelize independent tile loops with Cilk Plus work-stealing | Fast convolution implementation · Single convolution layer, 1024×1024 image, 4×4 kernels, 18 cores | multicore scalability: 95% | Helped |
| Synthesize transforms with D^N as an integer multiple of AVX width | Fast convolution implementation · Example square 2D and 3D kernels | algorithmic speed-up: 2D: 2.25, 2.25, 6.25, 6.25, 5.06; 3D: 3.38, 3.38, 15.63, 15.63, 11.39 | Helped |
| Increase channel and kernel counts to amortize transform costs | Fast tensor convolution · C3D kernels (3×3×3) | acceleration over direct convolution: greater than 6-fold | Helped |
| Use the fast tensor-convolution transforms rather than direct sliding-window convolution | C3D convolution · 3×3×3 kernels | theoretical speed-up: 8-fold | Helped |
| Use sparse transform matrices | Fast tensor convolution on CPU · C3D convolution | multiplications: 4M^2 + 24M -> 4M^2 + 15M | Helped |
| Use fast tensor convolution instead of Winograd's minimal algorithm | Convolution layers with M channels and M kernels · Layer convolution | multiplications: 4M^2 + 24M; 6M^2 | Mixed |
Closest papers by meaning
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.88)
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.88)
- InceptionNeXt: When Inception Meets ConvNeXt (CVPR 2024 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.86)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.86)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.85)
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.