How to make Matlab use Nvidia???

My computer has a CPU and a Nvidia Quadro. I tried script from "transfer learning with AlexNet" and it took me 1 h to finish it. And as the script was running it said that it was running on a single CPU. How can I make sure that Matlab uses the Nvidia that is installed on my computer????

I would really appreciate the help

Answer 1

[Walter Roberson]

https://www.mathworks.com/help/nnet/deep-learning-basics.html

"You can train a convolutional neural network on either a CPU, a GPU, or multiple GPUs and/or in parallel. Training on a GPU or in parallel requires the Parallel Computing Toolbox™. Using a GPU requires a CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher. Specify the training parameters including the execution environment using the trainingOptions function."

Some models of Quadro have compute capacity as low as 1.0; you did not happen to mention which one you use. You need to have the Parallel Computing toolbox, And you have to request that the GPU be used.

[Safia]

thanks for you reply. I don't remember exactly which Nvidia... However I think it was Nvidia Quadro FX 4??.... If this would help to answer the question?

I have the parallel computing toolbox, but how do I request that the GPU is used??

So lets say that Nvidia won't work for some reason, then my CPU will manage to conduct the transfer learning (however for a longer time), right??

[Walter Roberson]

See https://www.mathworks.com/help/nnet/ref/trainingoptions.html#input_argument_namevalue_d119e53360 for enabling GPU.

The Quaudro FX 4* were Compute Capacity 1.0, 1.1, or 1.3 -- none of which were enough for use in any but the very first release of the Parallel Computing Toolbox, if memory serves me. https://developer.nvidia.com/cuda-legacy-gpus; my checks show that Compute Capacity 1.3 was indeed the minimum that was ever supported for Parallel Computing Toolbox, as that was the first version that supported Double Precision. The Quadro FX 4800 and FX 4800 for Mac were the only two Quadro FX 4* to ever have been supported by Parallel Computing Toolbox in any release (and not for CNN work.)

[Safia]

I just checked it. It is a NVIDIA Quadro FX 580. The computer has 2GB RAM and 4 CPUs

I am really new to this field when it comes to the components of te computers.Just saw that FX 580 has a capacity of 1.1.

Do you recommend that I should buy an entire new computer or should I just buy additional RAM and a GeForce GTX 1060 6 GB?? Any suggestions?

[Walter Roberson]

The GTX 1060 card takes 120 Watts by itself and it is recommended that the system have a 400 W power supply. My suspicion would be that your current system might not be up to that.

The specs you list for your computer suggests it was built around 2010. An upgrade would be technically recommended -- but the costs would not be trivial.

https://www.nvidia.com/en-us/geforce/products/10series/geforce-gtx-1060/

[Walter Roberson]

Note that if you were doing a lot of this kind of work and if you cannot use single precision, and if you have enough money, then instead of the GTX 1060 you might want to get one of NVidia's Deep Learning Workstations, or equivalent system. The

There is an interesting analysis at http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/

Use of double precision is common in deep learning, but some systems can be trained well with single precision. Double precision speeds do not scale the way you would think they should: there is effectively double precision done in software at about 1/32 of single precision floating point speed, to which specialized double precision processors can be added in the design that might take you up to 1/8 of single precision floating point speed or even 1/3 of single precision floating point speed for the real $$$$ top end devices.

https://www.microway.com/knowledge-center-articles/comparison-of-nvidia-geforce-gpus-and-nvidia-tesla-gpus/