microTVM: TVM on bare-metal

microTVM runs TVM models on bare-metal (i.e. IoT) devices. microTVM depends only on the C standard library, and doesn’t require an operating system to execute. microTVM is currently under heavy development.

microTVM is:

• an extension to TVM’s compiler to allow it to target microcontrollers

• a way to run the TVM RPC server on-device, to allow autotuning

• a minimal C runtime that supports standalone model inference on bare metal devices.

Supported Hardware

microTVM currently tests against Cortex-M microcontrollers with the Zephyr RTOS; however, it is flexible and portable to other processors such as RISC-V and does not require Zephyr. The current demos run against QEMU and the following hardware:

• STM Nucleo-F746ZG

• STM STM32F746 Discovery

• nRF 5340 Development Kit

Getting Started with microTVM

Before working with microTVM, we recommend you have a supported development board. Then, follow these tutorials to get started with microTVM:

1. Start the microTVM Reference VM. The microTVM tutorials depend on Zephyr and on a compiler toolchain for your hardware. The reference VM is a convenient way to install those dependencies.

2. Try the microTVM with TFLite Tutorial.

3. Try running a more complex CIFAR10-CNN model.

How microTVM Works

You can read more about the design of these pieces at the microTVM Design Document.

Help and Discussion

The TVM Discuss Forum is a great place to collaborate on microTVM tasks, and maintains a searchable history of past problems.