I just started using Flux.jl. I have a loss function whose gradient is calculated much slower on GPU than on CPU. As a machine learning project typically involves a large dataset, it is difficult to create a MWE, but I wonder if someone have any suggestions by looking at the types of the arguments…

[image] wsshin: [9] generic_matmatmul!(C::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, tA::Char, tB::Char, A::Base.ReshapedArray{Float32, 2, PermutedDimsArray{Float32, 3, (2, 1, 3), (2, 1, 3), CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}}, Tuple{Base.MultiplicativeInverses.SignedMultiplicativeInver…

Gradient on GPU 70X slower than on CPU

Specific Domains Machine Learning

avikpal October 28, 2022, 4:08pm 8

Do you see any scalar indexing warning? You can disable it by CUDA.allowscalar(false)

4 Likes

Topic		Replies	Views
Code using Flux slow on GPU GPU flux	9	3115	November 6, 2019
Flux.jl: training fails at GPU but works on CPU Machine Learning gpu , flux	1	642	September 19, 2019
MNIST GPU CuArrays error GPU	23	3157	January 22, 2019
Flux: Scalar getindex error Machine Learning	13	2123	May 15, 2020
GPU gradient issues with matrix tranpose GPU flux	0	347	March 14, 2021

Gradient on GPU 70X slower than on CPU

Related topics