slow training on single gpu

hey, i'm trying to train inception v3 on single gpu. it takes about 21 hours for 20,000 iteration. it takes more than an hour for 1000 iteration of 32 images in a minibatch. caffe and tensorflow are 10 times faster on the same computer. in caffe it takes 7 minutes for 1000 iterations. how can i improve the training on matlab? Thanks

[Walter Roberson]

... install a faster GPU, perhaps with more memory?

There can be big performance differences between different GPUs, especially if double precision is being used. A higher GPU clock rate does not necessarily mean that it will be the best for double precision: some GPUs have special double precision units that speed processing up a lot.

[tomer cz]

as i wrote, the problem is when i train the net with matlab. with other frameworks it run up to 10 times faster on the same computer. so faster GPU is not the solution. it something with the setup of matlab. a friend check it on a computer with 3 gpus and it still run 2 times slower than 1 gpu with caffe.

Answer 1

[Joss Knight]

Upgrade MATLAB with each new release, we are making big performance improvements all the time.

[Chris P]

Only certain matalb versions can be used with particular CUDA toolkits though

[Joss Knight]

MATLAB has its own copies of the CUDA libraries, so the toolkit you install is irrelevant unless you are compiling your own CUDA code.

[Walter Roberson]

I think maybe the point is that newer CUDA toolkits do not support some of the older architectures, and newer MATLAB versions do not support older CUDA toolkits.

[Joss Knight]

The only dependency is the driver and the MATLAB version, since MATLAB carries the toolkit with it and it makes no difference what toolkit you install. Maybe that's what you're saying.