cuDNN: Efficient Primitives for Deep Learning
By Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan M. Cohen, John Tran, Bryan Catanzaro, Evan Shelhamer. 2014
1,991 citations recorded by the source corpus.
We present a library that provides optimized implementations for deep learning primitives. Deep learning workloads are computationally intensive, and optimizing the kernels of deep learning workloads is difficult and time-consuming. As parallel architectures evolve, kernels must be reoptimized for new processors, which makes maintaining codebases difficult over time. Similar issues have long been addressed in the HPC community by libraries such as the Basic Linear Algebra Subroutines (BLAS) [2]. However, there is no analogous library for deep learning. Without such a library, researchers implementing deep learning workloads on parallel processors must create and optimize their own implementations of the main computational kernels, and this work must be repeated as new parallel processors emerge. To address this problem, we have created a library similar in intent to BLAS, with optimized routines for deep learning workloads. Our implementation contains routines for GPUs, and similarly to the BLAS library, could be implemented for other platforms. The library is easy to integrate into existing frameworks, and provides optimized performance and memory usage. For example, integrating cuDNN into Caffe, a popular framework for convolutional networks, improves performance by 36% on a standard model while also reducing memory consumption.
Subfields: Computer Networks and Communications, Computer Vision and Pattern Recognition, Hardware and Architecture, Efficiency & Compression, ML Systems & Hardware
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Networks and Communications
- Yoshua Bengio
- Carole-Jean Wu
- Kyunghyun Cho
- Kaiming He
- Razvan Pascanu
- Aaron C. Courville
- Ian J. Goodfellow
- Pieter Abbeel
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Trevor Darrell
- Ming-Hsuan Yang
- Xiaogang Wang
- Jitendra Malik
- Andrew Zisserman
- Yi Yang
In Hardware and Architecture
- Yoshua Bengio
- Aaron C. Courville
- Torsten Hoefler
- Yann LeCun
- Kyunghyun Cho
- Kurt Keutzer
- Quoc V. Le
- Tal Ben-Nun
In Efficiency & Compression
In ML Systems & Hardware
Citations in the corpus
- Beyond human-level accuracy: computational challenges in deep learning (2019, background): “Better cache tiling, kernel optimization and fusion techniques might also help [10, 11].”
- PyTorch: An Imperative Style, High-Performance Deep Learning Library (2019, background): “Specialized libraries such as cuDNN [22], along with a body of academic work (such as [23] and [24]), produced a set of high-performance reusable deep learning kernels that enabled frameworks such as Caffe [1], Torch7 [25], or TensorFlow [3] to take advantage of these hardware accelerators.”
- TensorFlow: A system for large-scale machine learning (2016, neutral): “Since then, GPU vendors have innovated in their support for machine learning: NVIDIA’s cuDNN library [13] for GPU-based neural network training accelerates several popular image models by 2–4 × when using version R4 in place of R2 [15].”
- TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems (2016, positive): “For example, there are a number of optimized libraries for performing matrix multiplies on different devices, including BLAS [15] and cuBLAS [39], or GPU libraries for convolutional kernels for deep neural nets such as cuda-convnet [28] and cuDNN [9].”
- MegDet: A Large Mini-Batch Object Detector (2017, methodology): “Most existing deep learning frameworks utilize the BN implementation in cuDNN [4] that only provides a high-level API without permitting modification of internal statistics.”
- Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin (2015, background): “We take advantage of work in increasing individual GPU efficiency for low-level deep learning primitives [9].”
- Wav2Letter++: A Fast Open-source Speech Recognition System (2018, methodology): “For certain operations we extend the core ArrayFire CUDA back-end with more efficient cuDNN operations [7] such as convolutions and RNN routines, among others.”
- IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures (2018, methodology): “, 2017) such as preparing the next batch of data for the learner while still performing computation, compiling parts of the computational graph with XLA (a TensorFlow Just-In-Time compiler) and optimising the data format to get the maximum performance from the cuDNN framework (Chetlur et al., 2014).”
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution (2016, methodology): “Our implementation uses Torch [52] and cuDNN [53]; training takes roughly 4 hours on a single GTX Titan X GPU.”
- Theano: A Python framework for fast computation of mathematical expressions (2016, methodology): “Efficient CUDA primitives for neural networks are implemented in the cuDNN library [20], in particular convolutions, pooling, and their gradients.”
- Deep Speech: Scaling up end-to-end speech recognition (2014, methodology): “We use the cuDNN library [2] to implement this first layer of convolution efficiently.”
- SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size (2016, background): “That said, most of these libraries use the same underlying computational back-ends such as cuDNN (Chetlur et al., 2014) and MKL-DNN (Das et al.”
- Semantic Image Segmentation via Deep Parsing Network (2015, methodology): “We parallelize these operations using matrix multiplication on GPU as [4] did, b12 can be computed within 30ms.”
- ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design (2018, methodology): “The convolution library is CUDNN 7.0 [23].”
- Sequence-to-Sequence Speech Recognition with Time-Depth Separable Convolutions (2019, methodology): “2 by unrolling the computation and making a single call to an efficient CuDNN [18] implementation, which is much faster than calling U separate kernels.”
- Recent advances in convolutional neural networks (2015, methodology): “testing process, such as cuDNN [75] and fbfft [76].”
- Deep Clustered Convolutional Kernels (2015, background): “Copyright 2015 by the author(s). creasing data-sets (Deng et al., 2009) and powerful new computing platforms (Chetlur et al., 2014).”
- Using Rule-Based Labels for Weak Supervised Learning: A ChemNet for Transferable Chemical Property Prediction (2017, methodology): “ChemNet was trained using a Tensorow backend [1] with GPU acceleration using NVIDIA CuDNN libraries[6].”
- Fast ConvNets Using Group-Wise Brain Damage (2015, background): “This structure is motivated by the observation that the majority of current implementations of generalized convolutions (including the most efficient one at the time of submission) [9, 16, 26, 12, 45, 37] compute generalized convolutions by reducing them to matrix multiplications (this reduction is ”