=========================== NPU Profiler: Neuron Studio =========================== Overview ======== `Neuron Studio` is an Integrated Development Environment (IDE) and AI toolkit designed to help the developer develop and optimize AI models on MediaTek platforms. .. figure:: /_asset/ai-workflow-overview-step4.png :align: center :width: 80% .. raw:: html

Supported Platforms and OS -------------------------- * **Supported Platforms:** `Genio 720-EVK (MT8391)`, `Genio 520-EVK (MT8371)`. * **Supported OS:** `Android`. Profiling Capabilities ---------------------- The tool monitors the following performance metrics to help identify bottlenecks: * **NPU Frequency:** Displays the operational clock speed of the hardware engines during inference. * **NPU Loading:** Represents the utilization percentage, indicating the workload intensity on the NPU cores. * **DRAM Utilization:** Monitors the memory occupancy required for model tensors and intermediate buffers. * **DRAM Bandwidth:** (Not supported). To obtain memory resource information, the developer must contact a MediaTek representative to request the `MediaTek Evaluation Tool (MET)`. .. figure:: /_asset/neruon_studio_profile.png :alt: Profile Overview of Neuron Studio :align: center :width: 80% Profile Overview Installation ============ 1. **Download the Android NDK.** Obtain the NDK from the `official Android NDK download page `_. 2. **Obtain Neuron Studio.** Contact the MediaTek representative to request the `Neuron Studio` installation package. 3. **Configure Neuron Studio.** * Execute ``neuron-studio.exe``. * Configure the Android NDK path: Navigate to ``File`` -> ``Settings`` -> ``Android NDK``, click ``+New``, select the NDK folder, and click ``OK``. * Restart `Neuron Studio` to apply the changes. Profiling Workflow ================== 1. **Connect the device.** Connect the Genio platform to the host PC using a Type-C cable via the Android Debug Bridge (ADB). 2. **Enable root access.** .. code-block:: bash adb root 3. **Record the trace.** * In the `Neuron Studio` interface, navigate to ``Profile`` -> ``Record`` -> ``MediaTek Devices on ADB`` and select the target device (e.g., ``mtxxxx``). * Run the target AI workload, such as a benchmark or application. * Stop the recording to display the trace and save the results as a ``.json`` file. 4. **Analyze the results.** The developer can perform analysis using the `Neuron Studio` interface or external tools such as `Perfetto`.