--- title: "TargetDevice" language: "en" type: "Symbol" summary: "TargetDevice is an option for certain functions that specifies on which device the computation should be attempted." keywords: - gpu net training - gpu - gpu evaluation - V100 - Nvidia - Volta GPU - CUDA - OpenCL - graphics processor - coprocessor canonical_url: "https://reference.wolfram.com/language/ref/TargetDevice.html" source: "Wolfram Language Documentation" related_guides: - title: "Neural Network Construction & Properties" link: "https://reference.wolfram.com/language/guide/NeuralNetworkConstruction.en.md" related_workflows: - title: "Run Neural Network Training on GPUs" link: "https://reference.wolfram.com/language/workflow/RunNeuralNetworkTrainingOnGPUs.en.md" related_functions: - title: "NetTrain" link: "https://reference.wolfram.com/language/ref/NetTrain.en.md" - title: "NetChain" link: "https://reference.wolfram.com/language/ref/NetChain.en.md" - title: "NetGraph" link: "https://reference.wolfram.com/language/ref/NetGraph.en.md" - title: "NetExternalObject" link: "https://reference.wolfram.com/language/ref/NetExternalObject.en.md" - title: "Parallelize" link: "https://reference.wolfram.com/language/ref/Parallelize.en.md" - title: "NetEvaluationMode" link: "https://reference.wolfram.com/language/ref/NetEvaluationMode.en.md" related_tutorials: - title: "Neural Networks in the Wolfram Language" link: "https://reference.wolfram.com/language/tutorial/NeuralNetworksOverview.en.md" --- # TargetDevice ⚠ *Unsupported in Public Cloud* TargetDevice is an option for certain functions that specifies on which device the computation should be attempted. ## Details * Possible settings for ``TargetDevice`` change depending on the system and the available hardware. * General settings include: | | | | ----- | -------------------------------- | | "CPU" | use the CPU | | "GPU" | use available dedicated hardware | * The specification ``TargetDevice -> "GPU"`` has the following interpretation, depending on ``\$SystemID`` : | | | | ---------- | -------------------------------------------- | | "CoreML" | "MacOSX-ARM64" | | "CUDA" | "Windows-x86-64" or "Linux-x86-64" | | "DirectML" | "Windows-x86-64" if no CUDA card is detected | * Settings specific to system and available hardware include: | | | | ---------- | ------------------------------------- | | "CoreML" | Apple CoreML framework | | "CUDA" | Nvidia CUDA API | | "DirectML" | Microsoft Direct Machine Learning API | * Currently, the only settings supported by ``NetTrain`` are ``"CPU"`` and ``"CUDA"``. * ``TargetDevice -> {"GPU", n}`` allows a specific GPU to be used, where ``n`` is an integer between 1 and the number of GPUs available on your computer. * In general, a specific value of ``n`` might not identify the same GPU when different backends such as ``"CUDA"`` and ``"DirectML"`` are used. * ``TargetDevice -> {"GPU", All}`` specifies that all available GPUs should be used jointly. * ``TargetDevice -> {"GPU", {n1, n2, …}}`` specifies that a specific subset of the GPUs should be used jointly. * On Apple silicon machines, ``TargetDevice -> "CoreML"`` attempts to perform the computation using the Apple Neural Engine. It is not currently supported for ``NetTrain`` and ``NetMeasurements.`` » * On Windows machines, ``TargetDevice -> "DirectML"`` can perform the computation using an integrated or discrete GPU with DirectX support. It is not currently supported for ``NetTrain`` and ``NetMeasurements.`` * ``TargetDevice -> "CUDA"`` requires NVIDIA GPUs with compute capability 3.7 or 5.0 and higher. * In a fresh Wolfram Language installation on Linux and Windows machines, ``TargetDevice -> "GPU"`` and related GPU settings will automatically download additional libraries. Such a download can also trigger again in case updates are available. The downloads can also be started manually by running ``PacletInstall["MXNetResources"]``, ``PacletInstall["ONNXRuntimeResources"]`` and ``PacletInstall["CUDAResources"]``. ## Examples (10) ### Basic Examples (1) Train a net using the system's default GPU: ```wl In[1]:= trained = NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> "GPU"] Out[1]= [image] ``` Evaluate the resulting net using the system's default GPU: ```wl In[2]:= trained[[image], TargetDevice -> "GPU"] Out[2]= 8 ``` ### Scope (6) #### Inference (5) Evaluate a model on the CPU: ```wl In[1]:= NetModel["Wolfram ImageIdentify Net V1"][RandomImage[], TargetDevice -> "CPU"] Out[1]= Entity["Concept", "IgneousRock::9r5r5"] ``` --- Evaluate a model on the system's default GPU: ```wl In[1]:= NetModel["Wolfram ImageIdentify Net V1"][RandomImage[], TargetDevice -> "GPU"] Out[1]= Entity["Concept", "IgneousRock::9r5r5"] ``` --- Specify the use of CoreML (on macOS ARM64): ```wl In[1]:= NetModel["Wolfram ImageIdentify Net V1"][RandomImage[], TargetDevice -> "CoreML"] Out[1]= Entity["Concept", "IgneousRock::9r5r5"] ``` --- Specify the use of a DirectML-compatible card (on Windows x86-64): ```wl In[1]:= NetModel["Wolfram ImageIdentify Net V1"][RandomImage[], TargetDevice -> "DirectML"] Out[1]= Entity["Concept", "IgneousRock::9r5r5"] ``` --- Specify the use of a CUDA-compatible card (on Windows x86-64 or Linux x86-64): ```wl In[1]:= NetModel["Wolfram ImageIdentify Net V1"][RandomImage[], TargetDevice -> "CUDA"] Out[1]= Entity["Concept", "IgneousRock::9r5r5"] ``` #### Training (1) Train a net using a specified CUDA GPU: ```wl In[1]:= NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> {"GPU", 1}] Out[1]= [image] ``` Train a net using two specified GPUs, where each GPU receives a batch of 16 training examples per training iteration: ```wl In[2]:= NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> {"GPU", {1, 2}}, BatchSize -> 32] Out[2]= [image] ``` Train a net using all available GPUs: ```wl In[3]:= NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> {"GPU", All}, BatchSize -> 64] Out[3]= [image] ``` ### Possible Issues (3) If the system GPU is not supported, the operation will fail: ```wl In[1]:= n = 100000; trainingData = RandomReal[1, {n, 4}] -> RandomReal[1, {n, 4}]; net = NetChain[{8, Tanh, 2048, Tanh, 2048, Tanh, 4}, "Input" -> 4]; trained = NetTrain[net, trainingData, TargetDevice -> "GPU"] ``` NetTrain::badtrgdev: TargetDevice -> GPU could not be used, please ensure that you have a compatible NVIDIA graphics card and have installed the latest operating system drivers. ```wl Out[1]= $Failed ``` --- Some settings are only supported for model evaluation and will fail during training: ```wl In[1]:= NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> "CoreML"] ``` NetTrain::trgdevbad: Invalid setting TargetDevice -> CoreML. ```wl Out[1]= $Failed ``` --- There is currently no CUDA support on macOS: ```wl In[1]:= NetTrain[NetModel["LeNet"], "MNIST", TargetDevice -> "CUDA"] ``` NetTrain::trgdevunsupp: TargetDevice -> CUDA is not supported on MacOSX platforms. ```wl Out[1]= $Failed ``` ## See Also * [`NetTrain`](https://reference.wolfram.com/language/ref/NetTrain.en.md) * [`NetChain`](https://reference.wolfram.com/language/ref/NetChain.en.md) * [`NetGraph`](https://reference.wolfram.com/language/ref/NetGraph.en.md) * [`NetExternalObject`](https://reference.wolfram.com/language/ref/NetExternalObject.en.md) * [`Parallelize`](https://reference.wolfram.com/language/ref/Parallelize.en.md) * [`NetEvaluationMode`](https://reference.wolfram.com/language/ref/NetEvaluationMode.en.md) ## Tech Notes * [Neural Networks in the Wolfram Language](https://reference.wolfram.com/language/tutorial/NeuralNetworksOverview.en.md) ## Related Guides * [Neural Network Construction & Properties](https://reference.wolfram.com/language/guide/NeuralNetworkConstruction.en.md) ## Related Workflows * [Run Neural Network Training on GPUs](https://reference.wolfram.com/language/workflow/RunNeuralNetworkTrainingOnGPUs.en.md) ## History * [Introduced in 2016 (11.0)](https://reference.wolfram.com/language/guide/SummaryOfNewFeaturesIn110.en.md) \| [Updated in 2024 (14.1)](https://reference.wolfram.com/language/guide/SummaryOfNewFeaturesIn141.en.md)