Overview

The Arm CMSIS Ethos-U pack provides the low-level driver for Arm Ethos-U55, Ethos-U65, and Ethos-U85 neural processing units (NPUs). The driver connects an ML inference runtime to the NPU and supports initialization, neural-network command-stream execution, interrupt handling, and performance monitoring.

Before deployment, the Vela compiler optimizes a quantized ML model for the selected Ethos-U target. At runtime, the Cortex-M application and ML inference runtime use this pack's driver to submit the compiled command stream and model memory regions to the NPU.

Memory modes

Ethos-U systems can place model constants, the tensor arena, and temporary data in different memories. Vela uses the selected memory mode and the target's memory-performance configuration to optimize data movement and execution.

• SRAM-only mode keeps both constants and the tensor arena in SRAM for low access latency, but the complete model and arena must fit there.
• Shared-SRAM mode keeps the tensor arena in SRAM while read-only constants reside in Flash, MRAM, or DRAM, reducing SRAM usage.
• Dedicated-SRAM mode places constants and the tensor arena in larger memory and uses SRAM as a fast staging area for selected data.

See the General documentation for memory-mode concepts and the Integration documentation for linker placement, memory attributes, cache policy, and driver configuration.

Integration and tutorial

The Integration guide describes the end-to-end workflow for creating an ML application for your target hardware. It covers selecting the device configuration, creating a CMSIS solution project, compiling the model with Vela, placing the generated model regions in physical memory, configuring the driver, and validating the complete system.

The tutorial uses Keil Studio for VS Code to start from an example that matches the target's Ethos-U NPU. It shows how to obtain the resolved MLOps and Vela settings, generate an NPU-optimized model, build the application, and extend the solution with a board layer and configuration for the physical target hardware.

The pack includes three Test-Ethos-U CMSIS integration-test solutions. Each includes an ML model and FVP simulator configuration for initial validation. You can extend an example with other ML models and a configuration for the target hardware to validate the system integration. Select the example that matches the target NPU:

Related examples

• CMSIS-Ethos-Integration applies the integration workflow to a real-world device.
• CMSIS-Ethos-Zephyr is the Test-Ethos-U55 example ported to Zephyr.

Zephyr

Zephyr applications use the Ethos-U driver to execute Vela-compiled ML models. The Vela target and memory configuration must match the driver and Zephyr board integration. See the Zephyr chapter for details.

Features

• Single-variant core-driver components for Arm Ethos-U55, Ethos-U65, and Ethos-U85, plus a multi-variant component for run-time NPU selection.
• Synchronous and asynchronous command-stream execution.
• Configurable command-stream and model-region memory access.
• Interrupt handling and Performance Monitoring Unit (PMU) support.
• Per-family configuration and per-instance configuration for multi-variant systems.
• Platform hooks for cache maintenance, address translation, locking, and logging.
• Guidance for Vela configuration, driver use, system integration, and Zephyr deployment.

Links

• Documentation
• Zephyr
• Repository
• Issues