Suchergebnisse
Suchergebnisse:
Cliff Woolley is a senior developer technology engineer with NVIDIA. He received his master's degree in Computer Science from the University of Virginia in 2003, where he was among the earliest academic researchers to explore the use of GPUs for general purpose computing.
View Cliff Woolley’s profile on LinkedIn, a professional community of 1 billion members. Experience: NVIDIA · Location: San Jose, California, United States · 12 connections on LinkedIn.
- 12
- 25
- NVIDIA
- San Jose, California, United States
1. Dez. 2014 · Cliff Woolley (NVIDIA) Philippe Vandermersch (NVIDIA) Jonathan Cohen (NVIDIA) John Tran (NVIDIA) Bryan Catanzaro (Baidu) Evan Shelhamer (UC Berkeley) Publication Date. Monday, December 1, 2014. Published in. Deep Learning and Representation Learning Workshop (NIPS2014) Research Area. Artificial Intelligence and Machine Learning. External Links.
- Abstract
- 2 Library
- 2.1 Overview and Handles
- 2.1.1 Spatial Convolutions
- 3.1 Our approach
- 4 Caffe Integration
- 4.1 Development
- 4.2 User Experience
- 6 Future Work
- 7 Conclusion
We present a library of efficient implementations of deep learning primitives. Deep learning workloads are computationally intensive, and optimizing their ker-nels is difficult and time-consuming. As parallel architectures evolve, kernels must be reoptimized, which makes maintaining codebases difficult over time. Similar issues have long been addre...
One of the primary goals of cuDNN is to enable the community of neural network frameworks to benefit equally from its APIs. Accordingly, users of cuDNN are not required to adopt any particular software framework, or even data layout. Rather than providing a layer abstraction, we provide lower-level computational primitives, in order to simplify int...
The library exposes a host-callable C language API, but requires that input and output data be resi-dent on the GPU, analogously to cuBLAS. The library is thread-safe and its routines can be called from different host threads. Convolutional routines for the forward and backward passes use a com-mon descriptor that encapsulates the attributes of the...
The most important computational primitive in convolutional neural networks is a special form of batched convolution. The parameters governing this convolution are listed in table 1. In this section, we describe the forward form of this convolution - the other forms necessary for backpropagation are closely related. There are two inputs to the conv...
NVIDIA provides a matrix multiplication routine that achieves a substantial fraction of floating-point throughput on GPUs. The algorithm for this routine is similar to the algorithm described in [20]. Fixed sized submatrices of the input matrices A and B are successively read into on-chip memory and are then used to compute a submatrix of the outpu...
Caffe is a deep learning framework developed with expression, speed, and modularity in mind. Its architecture mirrors the natural modularity of deep networks as compositional models made from a collection of inter-connected layers. A deep network is defined layer-by-layer in a plaintext schema. Each layer type is implemented according to a simple p...
Integration is made simple by the self-contained design of the cuDNN handles, descriptors, and function calls together with the modularity of the framework. The core Caffe framework is unal-tered, preserving the network, layer, and memory interfaces. Changes were isolated to new layer definitions and implementations, helper functions for descriptor...
cuDNN computation is transparent to the user through drop-in integration. The model schema and framework interfaces are completely unchanged. Setting a single compilation flag during installation equips Caffe with cuDNN layer implementations and sets cuDNN as the default computation engine. Layers automatically fall back to the standard Caffe funct...
We are considering several avenues for expanding the performance and functionality of cuDNN. Firstly, although our convolution routines are competitive with other available implementations, more work remains to bring performance up to that attained by matrix multiplication. Over time, we hope to shrink this gap. Secondly, we envision adding support...
This paper presents cuDNN, a library for deep learning primitives. We presented a novel implemen-tation of convolutions that provides reliable performance across a wide range of input sizes, and takes advantage of highly-optimized matrix multiplication routines to provide high performance, without requiring any auxiliary memory. We also provide a s...
3. Okt. 2014 · cuDNN: Efficient Primitives for Deep Learning. Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, Evan Shelhamer. We present a library of efficient implementations of deep learning primitives.
- Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, Ev...
- arXiv:1410.0759 [cs.NE]
- 2014
Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, Evan Shelhamer: cuDNN: Efficient Primitives for Deep Learning. CoRR abs/1410.0759 (2014) 2000 – 2009. see FAQ. What is the meaning of the colors in the pu ...
Sylvain Jeaugey, NVIDIA | Cliff Woolley, NVIDIA | Sreeram Potluri, NVIDIA | Ke Wen, NVIDIA | Nathan Luehr, NVIDIA GTC 2020 NCCL (NVIDIA Collective Communication Library) optimizes inter-GPU communication on PCI, NVIDIA NVLink, and Infiniband, powering large-scale training for most DL frameworks, including Tensorflow, PyTorch, MXNet, and Chainer.
