Ah, this is the problem. We are making use of cuDNN, which requires compute capability >=3.0 (https://developer.nvidia.com/cudnn).

It is a bug though that we are not catching this properly and displaying a useful message.

@Marco Thiel, @David Proffer, @Arno Bosse: We are looking into a solution to this. I will have an update for you within a week.

When I run the example.nb above, everything works fine. I am assuming my GTX 660 is compatible but TargetDevice is just not a parameter of Classify yet.

This should be added because everything I have built uses Classify or Predict and forcing users like myself to specify the parameters of super complicated NeuralNetworks in order to utilize GPUs is not going to be good for mass adoption.

BTW: It only took about 2 minutes to finish the full training in the example.nb which is amazing!

Can you post your full example that you think should be working? Its impossible to debug this without more information.

Again, can you post the code that is failing for your example?

Also, what is the failure message with your "I first tried the following example available on the Wolfram website" example?

The documentation is doing something slightly confusing. It is reassigning to the value "lenet".

Try the following. Rename "lenet" to "trainedlenet" after it has been trained. You will want to re-run everything from the top after doing this:

lenet = NetChain[{ConvolutionLayer[20, 5], Ramp, PoolingLayer[2, 2], 
   ConvolutionLayer[50, 5], Ramp, PoolingLayer[2, 2], FlattenLayer[], 
   500, Ramp, 10, SoftmaxLayer[]}, 
  "Output" -> NetDecoder[{"Class", Range[0, 9]}], 
  "Input" -> NetEncoder[{"Image", {28, 28}, "Grayscale"}]];

trainedlenet = NetTrain[lenet, trainingData, ValidationSet -> testData, 
  MaxTrainingRounds -> 3, TargetDevice -> "GPU"]

This way you can be sure to not get them mixed up.

In your picture, some of the code is black and some of the code is blue.

The syntax highlighter highlights things in blue if they are undefined. This means that it thinks "lenet", "trainingdata" and "testData" are all undefined for some reason.

I've attached my notebook. It shows a simple evaluation which leads to a failure (because I don't have a compatible graphics card). But you can see it runs.

Attachments:

There does seem to be problem. It probably only happens with the your graphics card or driver. We haven't been able yet to reproduce it.

What Operating System are you using? Additionally, do you know what your graphics card is?

We would like to reproduce the problem as closely as possible. If you can, please send an email to support@wolfram.com about this issue. If you can, include a notebook where you have evaluated the command SystemInformation[]. This will tell us more about your computer and its graphics card, helping us to fix this issue.

I am running Mathematica on Ubuntu 16.04. Here is the same info for my system:

Perhaps it is related to the CUDA Compute Capabilities of my Geforce GTX 460.

Gijsbert

Here is my info:

In[9]:= CUDAInformation[]

Out[9]= {1 -> {"Name" -> "GeForce GTX 660", "Clock Rate" -> 888500, 
   "Compute Capabilities" -> 3., "GPU Overlap" -> 1, 
   "Maximum Block Dimensions" -> {1024, 1024, 64}, 
   "Maximum Grid Dimensions" -> {2147483647, 65535, 65535}, 
   "Maximum Threads Per Block" -> 1024, 
   "Maximum Shared Memory Per Block" -> 49152, 
   "Total Constant Memory" -> 65536, "Warp Size" -> 32, 
   "Maximum Pitch" -> 2147483647, 
   "Maximum Registers Per Block" -> 65536, "Texture Alignment" -> 512,
    "Multiprocessor Count" -> 6, "Core Count" -> 192, 
   "Execution Timeout" -> 1, "Integrated" -> False, 
   "Can Map Host Memory" -> True, "Compute Mode" -> "Default", 
   "Texture1D Width" -> 65536, "Texture2D Width" -> 65536, 
   "Texture2D Height" -> 65536, "Texture3D Width" -> 4096, 
   "Texture3D Height" -> 4096, "Texture3D Depth" -> 4096, 
   "Texture2D Array Width" -> 16384, 
   "Texture2D Array Height" -> 16384, 
   "Texture2D Array Slices" -> 2048, "Surface Alignment" -> 512, 
   "Concurrent Kernels" -> True, "ECC Enabled" -> False, 
   "TCC Enabled" -> False, "Total Memory" -> 1610612736}}

In[10]:= CUDAResourcesInformation[]

Out[10]= {{"Name" -> "CUDAResources", "Version" -> "10.5.0", 
  "BuildNumber" -> "", "Qualifier" -> "Win64", 
  "WolframVersion" -> "10.5+", "SystemID" -> {"Windows-x86-64"}, 
  "Description" -> "{ToolkitVersion -> 7.0, MinimumDriver -> 300.0}", 
  "Category" -> "", "Creator" -> "", "Publisher" -> "", 
  "Support" -> "", "Internal" -> False, 
  "Location" -> 
   "C:\\Users\\David\\AppData\\Roaming\\Mathematica\\Paclets\\\
Repository\\CUDAResources-Win64-10.5.0", "Context" -> {}, 
  "Enabled" -> True, "Loading" -> Manual, 
  "Hash" -> "79fa747a52a45bf2d78e2c3516c80061"}}

No, Classify has no support for GPU (as you can see from docs).

This is a known issue, see:

http://community.wolfram.com/groups/-/m/t/917616

GW

This issue seems to be resolved in version 11.1 (at least for my GTX 1050)!

Gijsbert

It's working great for NetChain on my Nvidia 1080. However, doesn't look like they built GPU support for Classify and Predict with

Classify[trainingData, {Method -> "NeuralNetwork", 
  PerformanceGoal -> "Quality", TargetDevice -> "GPU"}]

Wish it did though! Put that on the list.

Seems that the GPU ability of NeuralNetwork is not supported in 11.1, this works fine on 11.0 on OS X

I suppose you are running a custom Linux machine rather than a normal out of the box Mac. I guess that is one way to get a GPU on Mac OS but it won't solve your CUDA SDK problem. Probably something you should disclose rather than making it seem like anyone who thinks GPU support wouldn't normally be available on Mac's is an idiot.

That is very cool that you had it working on 11.0. Congrats on that. Last summer, I was not able to use GPU on either my iMac nor my Macbook Pro. Wolfram Support told me that most Mac's don't generally have GPU's so I figured it was a general truth. I found a couple external GPU options but none were really more cost effective over just buying a new Alienware with 1080 GPU, which I did.

Another cloud option everyone can use is this.

Running Mathematica on the Cloud with Rescale.

Video: https://support.rescale.com/customer/en/portal/articles/2651932-running-mathematica-on-the-cloud-with-rescale

Does this mean that OSX users who previously used nVidia cards with an eGPU over Thunderbolt (e.g. Akitio Node or BizonBox) will no longer be able to do so in v11.1? I was planning on purchasing such a set-up myself (c. $550 for Akitio Node and a used GeForce 980Ti 6GB). Update: Nevermind, I see now from the context that the answer is likely 'yes' – this won't work anymore.

Oh, what an unfortunate choice! I just bought 4 BizonBoxes at the beginning of this year only because of the Neural Network framework of Mathematica. It worked really nicely for MMA 11.0.1 on OSX and opened up the large and powerful new expansion of the Wolfram Language. This decisions cuts all OSX users off a large chunk of the new development in the Wolfram Language.

I also do not think that it is just the few old OSX legacy devices with GPUs that you stop supporting, but it is cutting off OSX users from using other techniques to access the neural network framework. It appears that much of the development of the Wolfram Language goes in the direction of machine learning. GPUs are crucial for that. Training times on my networks went down from a day to a couple of minutes.

I believe that there were several solutions available so that OSX users could make use of this exciting new development in the Wolfram Language, but this decision seems to cut all of us off, which is unfortunate I think.

Best wishes,

Marco

Just having updated to MM 11.1 and finding out that GPU training on Mac OS X fails without any announcement or description in the Documentation of Nettrain[]. As for other users in the forum also for our organization this makes it difficult for us to use Mathematica as training platform for MachineLearning. In our algorithm development department, we have mainly 15" MacBook Pro Mid/Late 2013 notebooks deployed, all with GPUs, also for CUDA/OpenCL Programming within Mathematica (still have to check that CUDALink and OpenCLLink both function in MM 11.1). A consistent GPU support across all platforms, including CUDA where available and OpenCL (works fine with many core CPUs or AMD GPUs) is a great plus for Mathematica.

Dear Sebastian,

thank you very much for that helpful reply. I will try to find some time to post some easy instructions of how to make the BizonBox work with Mathematica and show some benchmarks.

Thanks a lot,

Marco

EDIT: Sorry for that. I have fixed the thing. Actually, it works better now than in 10.4. I did not need to perform one step and that caused trouble now. The thing works out of the box.

FYI... April 4, 2017 - Apple pushes the reset button on the Mac Pro https://techcrunch.com/2017/04/04/apple-pushes-the-reset-button-on-the-mac-pro/

I ask him what Apple’s philosophy on external GPUs is. (Matthew Panzarino (@panzer) ) “I think they have a place,” he says, and leaves it at that. (Craig Federighi)

Dear Sebastian,

it is great to see that the GPU framework is enabled for Mac users again. I was just wondering one thing. As I was preparing a post with instructions of how to use the Bizon Box with Mathematica I tried to run some old code. It appears that the option "CompilerInstallation" is not valid anymore. Is that by design?

Best wishes,

Marco

Given the Mac Pro and the latest powerful MacBook pros, etc. Wolfram should implement Metal support so that we can get GPU acceleration even on modern hardware without NVIDIA GPUs... Any words on this from Wolfram? They also advertise OpenCL compatibility, that framework was an open source effort from Apple to abstract away CUDA so that it would also work on non-NVIDIA gpus.. I guess it's Metal now. The documentation at https://reference.wolfram.com/language/OpenCLLink/tutorial/Setup.html#10897850 looks outdated. Given that Apple will likely add/switch to ARM processors even for Macs in a year or two, I hope Wolfram gets up to speed on adopting Metal asap. On top of that the processor in the iPad Pro is already more powerful than some mac notebooks (GPUs included).