Abstract
This guide provides tips for improving the performance of convolutional layers, with details on the impact of parameters including batch size, input and filter dimensions, stride, and dilation.
1. QuickStart Checklist
The following quickstart checklist provides specific tips for convolutional layers.

Choose the number of input and output channels to be divisible by 8 to enable Tensor Cores. For the first convolutional layer in most CNNs where the input tensor consists of 3channel images, padding to 4 channels is sufficient if a stride of 2 is used; see Channels In And Out.

Choose parameters (batch size, number of input and output channels) to be divisible by at least 64 and ideally 256 to enable efficient tiling and reduce overhead; see Quantization Effects.

Larger values for sizerelated parameters (batch size, input and output height and width, and the number of input and output channels) can improve parallelization. As with fullyconnected layers, this speeds up an operation’s efficiency, but does not reduce its absolute duration; see How Convolution Parameters Affect Performance and subsections.

NVIDIA libraries offer a set of different convolution algorithms with different performance behaviors, dependent on the convolution’s parameters. When the size of the input processed by the network is the same in each iteration, autotuning is an efficient method to ensure the selection of the ideal algorithm for each convolution in the network. For TensorFlow, autotuning is enabled by default. For PyTorch, enable autotuning by adding torch.backends.cudnn.benchmark = True to your code.

Choose tensor layouts in memory to avoid transposing input and output data. There are two major conventions, each named for the order of dimensions: NHWC and NCHW. We recommend using the NHWC format where possible. This is supported natively in MXNet and available through XLA in TensorFlow; native support for PyTorch is in development. Additional details can be found in Tensor Layouts In Memory: NCHW vs NHWC.
2. Introduction
A convolution is defined by the sizes of the input and filter tensors and the behavior of the convolution, such as the padding type used.
Figure 1 illustrates the minimum parameter set required to define a convolution.
Parameter  Tensor  Meaning 

N  N/A  Batch size 
C  Input  Number of channels 
H  Height  
W  Width  
K  Output  Number of channels 
P  Height (often derived from other parameters)  
Q  Width (often derived from other parameters)  
R  Filter  Height 
S  Width  
U  Vertical stride  
V  Horizontal stride  
PadH  Input padding in the vertical dimension  
PadW  Input padding in the horizontal dimension  
DilH  Dilation in the vertical dimension  
DilW  Dilation in the horizontal dimension 
3. Convolution Algorithms
NVIDIA cuDNN library implements convolutions using two primary methods: implicitGEMMbased and transformbased.
The implicit GEMM approach is a variant of direct convolution, and operates directly on the input weight and activation tensors. Alternatively, convolutions can be computed by transforming data and weights into another space, performing simpler operations (for example, pointwise multiplies), and then transforming back. The cuDNN library provides some convolution implementations using FFT and Winograd transforms.
3.1. Choosing A Convolution Algorithm With cuDNN
When running a convolution with cuDNN, for example with cudnnConvolutionForward, you may specify which general algorithm is used.
The cuDNN API provides functions for estimating the relative performance of different algorithms. One set of functions, prefixed with cudnnGet, uses a set of heuristics to predict the relative performance of available algorithms. These functions evaluate quickly; however, although we’re constantly improving our heuristics, the predictions may not always be accurate. Suboptimal algorithm choice may occasionally occur, and is more common for unusual types of convolutions and corner cases.
An alternative set of functions, prefixed with cudnnFind, tests and reports the performance of all available algorithms to determine the most efficient option for the given convolution operation. The benefit of using these functions is that the algorithm selected is the best choice. However, since actual performance tests are being run, these functions can be time and resourceintensive.
After an algorithm is chosen, our heuristics specify additional lowlevel details; for example, tile size, discussed at length in Quantization Effects and in the Dimension Quantization Effects section in the Matrix Multiplication Background User Guide. These parameters are not exposed in the API.
4. Tensor Core Usage And Performance Recommendations
The primary method to execute convolutions (without transforms) used by NVIDIA Tensor Core GPUs is called implicit GEMM. It performs exactly the same number of math operations as a direct convolution and hence is computationally equivalent.
Implicit GEMM operates natively on the convolution input tensors, converting the computation into a matrix multiply on the fly. It is important to note that corresponding matrices are never created in memory. Thus, to calculate arithmetic intensity, one can use the original tensor sizes.
To illustrate the concept of convolution as a matrix multiply let’s first consider a single application of a convolution filter to input data. Say we are applying a 3x3 convolution to a 128channel input tensor. To compute a single output value, we effectively compute a dotproduct of two 1,152element vectors. One is the weights for a filter (3x3x128 = 1,152). The other is composed of the data (activation) values that are multiplied with the weights to produce the output. Since not all 1,152 data values are contiguous in memory, the original tensor layout is read and on the fly is converted to the appropriate vector form. To compute all the outputs we perform multiple dotproducts, which can be seen as a matrix multiply, but since the matrices are implicitly formed, rather than created in memory, this method is called implicit GEMM. To understand the performance of convolutions, however, it can be useful to understand the shapes and sizes of these “virtual” matrices.
4.1. Tensor Layouts In Memory: NCHW vs NHWC
Convolutions typically operate on fourdimensional tensors: a batch composed of N “images” of C channels of H x W feature maps.
In practice, NHWC layouts are natively available and well supported in MxNet, and can be used via XLA in TensorFlow. Native PyTorch implementations are in development. Performance examples in this section can be assumed to use input and output data in the NHWC layout unless otherwise stated.
4.2. Implicit GEMM Dimensions
Let’s now consider the dimensions of the matrices we encounter when performing forward convolution, calculating activation gradients, and calculating weight gradients.
Computation Phase  "M"  "N"  "K" 

Forward Propagation  N*P*Q  K  C*R*S 
Activation Gradient  N*H*W  C  K*R*S 
Weight Gradient  C*R*S  K  N*P*Q 
The composition of the “virtual” matrices is shown in Figure 3. For each pass, there is one virtual matrix that, if explicitly constructed, would contain more values than its corresponding tensor. For example, during forward convolution, the A matrix (N*P*Q x C*R*S) is composed of input activations (a tensor with dimensions N x H x W x C). Each individual input activation appears in R*S places in the matrix, repeated with necessary offsets to cause multiplication of that input value with the overlaid values of the matching R x S filter channel. Similar conceptual expansions occur for output activation gradients when computing input activation gradients, and for input activations during weight gradient calculation.
For example, computing a 3x3 convolution on a 256x56x56x64 input tensor, producing a 256x56x56x128 output, all in halfprecision, has an arithmetic intensity of 383.8 FLOPS/byte.
4.3. Quantization Effects
Tile and wave quantization effects can be significant, especially for small problem sizes (background in Dimension Quantization Effects in the Matrix Multiplication Background User Guide). Just like for GEMMs, in implicit GEMM, the representation of the output matrix is divided into tiles of a chosen size, and that set of tiles is distributed across available multiprocessors.
Our testing GPU has 80 SMs. Each SM can handle a number of thread blocks in parallel that is dependent on the kernel being used; for best parallelization, an implicit GEMM should contain an integer multiple of 80 tiles.
As Figure 4 shows, convolutions that result in small equivalent GEMMs can exhibit significant quantization effects. When N = 40, 160 tiles are created; when N = 41, 164 tiles are created. The former results in high Tensor Core utilization, while the latter will require an additional wave to process the remainder of 4 tiles, severely impacting performance (Figure 4 (a)). Once the convolutions are reasonably large, the effect is less pronounced (Figure 4 (b)).
It is worth noting that weight gradient quantization does not behave as implied by the GEMM dimensions in Figure 3; for quantization purposes, the height of the matrix to be tiled should be seen as C (rather than C*R*S). This is discussed in more detail at the end of the section on filter size.
4.4. How Convolution Parameters Affect Performance
In this section, we discuss the trends affecting performance. To keep things simple, padding is set such that H = P and W = Q, and both the stride and dilation are equal to one unless indicated otherwise.
4.4.1. Batch Size, Height And Width
When representing a forward convolution as a GEMM, the product of batch size, output height, and output width (N*P*Q) is the “M” dimension of the unrolled input tensor (A matrix) as well as the output (C matrix).
The individual values of these parameters are not especially important to GEMM performance; only the final dimension, the product N*P*Q, is significant. Conveniently, in most applications, batch size may be changed more easily than the parameters contributing to the height and width of the output tensor.
Generally, efficiency improves as N*P*Q increases, with diminishing returns. From Figure 5, we can see that points with equivalent N*P*Q have roughly equivalent performance, as the corresponding GEMMs have the same dimensions.
When calculating the activation gradient, N*H*W is the “M” dimension of the equivalent GEMM (compared to N*P*Q in forward convolution). With a filter stride of 1, the performance of forward convolution and activation gradient calculation will be roughly the same.
In contrast, for weight gradient calculation, N*P*Q becomes the accumulation (“K” in a GEMM) dimension. The performance impact of this dimension is not as straightforward, as it does not affect the tiling of the output matrix in any way. However, larger values of N*P*Q generally leads to more time spent multiplying and accumulating elements, rather than in setup and teardown overhead for the GEMM computation, improving the fraction of peak performance achieved for the whole operation. It is worth noting that cuDNN generally supports tiling in the N*P*Q dimension as well for weight gradients, as many common layers result in tiny (for example, 64x64 in the first block of a standard ResNet) output matrices that, by themselves, don’t offer enough tile parallelism to keep the GPU occupied.
4.4.2. Filter Size
The equivalent GEMM for forward convolution has a “K” dimension of C*R*S. As mentioned previously, the “K” dimension does have an impact on performance, and this effect is most pronounced for small GEMMs.
When using a 1x1 filter, layers with more input channels tend to perform better in forward convolution (Figure 6), as it is ultimately the C*R*S product that matters.
When calculating activation gradients of the convolution, K affects this dimension instead: “K” = K*R*S. There is a clear similarity between Figure 6 and Figure 7; in general, trends related to C for forward convolution are related to K for activation gradient calculation and vice versa.
The weight gradient calculation has “M” = C*R*S, therefore the impact of filter size on performance is similar to that discussed for batch size, height, and width previously; larger values tend to perform better. When considering tile quantization, however, the weight gradient algorithm differs from the forward and data gradient ones; only the C dimension, not the full C*R*S dimension, quantized against the tile size (meaning, filter size parameters can be ignored). For example, when using 64x64 tiles, if C = 32, half of each tile (vertically) is wasted regardless of R and S; only the value of C matters.
4.4.3. Channels In And Out
Using Tensor Cores on Volta requires that C and K be multiples of 8 in FP16 or 16 in INT8. A caveat applies here: currently, for NCHWpacked FP16 data, channels will be automatically padded to multiples of 8 such that tensor cores will be enabled.
However, using NCHW data with tensor core enabled kernels involves some additional transpose costs, which are discussed in Tensor Layouts In Memory: NCHW vs NHWC. On the other hand, automatic padding doesn’t kick in for NHWCpacked data, so a lessefficient fallback kernel, which does not make use of tensor cores, is chosen. Convolutions with NHWC data do perform better than those with NCHW data given that C and K are divisible by 8. In other words, if a layer is already being used with NCHW data, automatic padding will occur; however, if NHWC data is being used, choosing or padding C and K to be a multiple of 8 improves performance.
Unfortunately, especially for layers near the beginning and end of a network, C and K can be small and nonnegotiable. For the first layer in a network, it is common to have a very small value of C (1 or 3 for grayscale and RGB or YCrCb images, respectively). Specialcase convolution implementations are available to meet this need, specifically for C = 4 and a stride of 2 (Figure 8), a common case in many CNNs.
Forward convolution performance relative to C was previously discussed in Filter Size (“K” = C*R*S), as was activation gradient calculation relative to K (“K” = K*R*S). The effect of C on weight update performance (with “M” = C*R*S) is mentioned in Batch Size, Height And Width. In short, larger values usually give higher efficiency, with diminishing returns.
The number of channels of input and output can have a more immediate impact on performance, however; the GEMM dimension “N” is equal to C or K for all of the forward convolution, activation gradient calculation, and weight gradient calculation. Thus, your choosing of these parameters can have a direct impact on performance.
Similar behavior can be seen in forward convolution and weight gradient computation as varied across K (Figure 9). In both cases, “N” = K, resulting in a strong trend for small values of K and diminishing returns once K is larger than most tile sizes.
The same effect is present for channels of input in activation gradient computation (“N” = C), as seen in Figure 10. As previously mentioned, the effect of C on activation gradient computation tends to match the effect of K on forward convolution.
4.5. Strides
Filter strides (U and V) impact performance mostly via their effect on input and output tensor dimensions. Using a horizontal and vertical stride of 1, H and W are roughly equal to P and Q respectively, depending on the filter size and padding. However, when larger strides are used, there is a severalfold difference in the size of input and output feature maps. In turn, this impacts GEMM dimensions and performance.
4.6. HighPerformance Example
An example of a convolution with high performance is shown in Figure 13. This scenario is based on a convolutional layer with input feature maps of size 32x32, filters of size 5x5, 1024 input channels, and 1024 output channels; each parameter is reasonably large. NHWC data is used to avoid overhead from additional transposes. Input and output channel numbers are divisible by 8, so Tensor Cores will be enabled.
With batch sizes of 32 or 64, performance for this case exceeds 100 TFLOPS for activation and weight gradient calculation. Forward convolution performs at just over 90 TFLOPS.
5. Convolution Variants
5.1. Dilated
Dilated convolutions are a variant of a regular convolutional layer that effectively expands the filter being applied by inserting zeros between filter elements.
The dilation factor is one greater than the number of zeros added between each pair of elements. As a result, the overall 2D area overlapping with each channel of the filter increases.
Choice of dilation factor affects how a convolution is represented as a virtual GEMM, but does not actually change the dimensions of that GEMM; therefore, performance for forward convolution and activation gradient computation is similar regardless of dilation factor (Figure 15).
Notices
Notice
_{This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material (defined below), code, or functionality.}
_{NVIDIA reserves the right to make corrections, modifications, enhancements, improvements, and any other changes to this document, at any time without notice.}
_{Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.}
_{NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed either directly or indirectly by this document.}
_{NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.}
_{NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document, ensure the product is suitable and fit for the application planned by customer, and perform the necessary testing for the application in order to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional or different conditions and/or requirements beyond those contained in this document. NVIDIA accepts no liability related to any default, damage, costs, or problem which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document or (ii) customer product designs.}
_{No license, either expressed or implied, is granted under any NVIDIA patent right, copyright, or other NVIDIA intellectual property right under this document. Information published by NVIDIA regarding thirdparty products or services does not constitute a license from NVIDIA to use such products or services or a warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual property rights of the third party, or a license from NVIDIA under the patents or other intellectual property rights of NVIDIA.}
_{Reproduction of information in this document is permissible only if approved in advance by NVIDIA in writing, reproduced without alteration and in full compliance with all applicable export laws and regulations, and accompanied by all associated conditions, limitations, and notices.}
_{THIS DOCUMENT AND ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. TO THE EXTENT NOT PROHIBITED BY LAW, IN NO EVENT WILL NVIDIA BE LIABLE FOR ANY DAMAGES, INCLUDING WITHOUT LIMITATION ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, HOWEVER CAUSED AND REGARDLESS OF THE THEORY OF LIABILITY, ARISING OUT OF ANY USE OF THIS DOCUMENT, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the products described herein shall be limited in accordance with the Terms of Sale for the product.}
VESA DisplayPort
_{DisplayPort and DisplayPort Compliance Logo, DisplayPort Compliance Logo for Dualmode Sources, and DisplayPort Compliance Logo for Active Cables are trademarks owned by the Video Electronics Standards Association in the United States and other countries.}
HDMI
_{HDMI, the HDMI logo, and HighDefinition Multimedia Interface are trademarks or registered trademarks of HDMI Licensing LLC.}
Trademarks
_{NVIDIA, the NVIDIA logo, and CUDA, CUDA Toolkit, GPU, NVLink, NVIDIA Deep Learning SDK, NVIDIA Developer Program, NVIDIA GPU Cloud, Turing, and Volta are trademarks and/or registered trademarks of NVIDIA Corporation in the United States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.}