search

hughperkins / DeepCL

OpenCL library to train deep convolutional neural networks
Mozilla Public License 2.0
865 stars 199 forks source link

does the Neural network created on GPU? #146

Open nano1900 opened 6 years ago

nano1900 commented 6 years ago

what i mean is ,say you created a network and when you traning it, does the network created on GPU(during a forward pass and backward pass all just runing on the GPU , GPU device do not communicate with host)or the network created on the host and used the GPU as just a compter device

hughperkins commented 6 years ago

The second. Cpu sends commands to the gpu. The data stays on the gpu in between layers, but the command to run the next layer comes from the cpu. Nuance: the program to run for each layer is called a "kernel", and is sent to the gpu driver only once.

On Fri, Mar 30, 2018, 05:55 nano1900 notifications@github.com wrote:

what i mean is ,say you created a network and when you traning it, does the network created on GPU(during a forward pass and backward pass all just runing on the GPU , GPU device do not communicate with host)or the network created on the host and used the GPU as just a compter device

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/hughperkins/DeepCL/issues/146, or mute the thread https://github.com/notifications/unsubscribe-auth/AAHiqPzThZxK_k1hCqvjDFYlmO1sTawYks5tjgEFgaJpZM4TBhaU .

nano1900 commented 6 years ago

oh! i see.The data(between layers) stay on the GPU , Host cpu send "kernels" to the GPU and the GPU operate these data

Thank you!

hughperkins commented 6 years ago

:)