Plumerai icon

Monday, October 4, 2021

The world’s fastest deep learning inference software for Arm Cortex-M

At Plumerai we enable our customers to perform increasingly complex AI tasks on tiny embedded hardware. We’re proud to announce that our inference software for Arm Cortex-M microcontrollers is the fastest and most memory-efficient in the world, for both Binarized Neural Networks and for 8-bit deep learning models. Our inference software is an essential component of our solution, since it directs resource management akin to an operating system. It has 40% lower latency and requires 49% less RAM than TensorFlow Lite for Microcontrollers with Arm’s CMSIS-NN kernels while retaining the same accuracy. It also outperforms any other deep learning inference software for Arm Cortex-M:

Inference time RAM usage
TensorFlow Lite for Microcontrollers 2.5 (with CMSIS-NN) 129 ms 155 KiB
Edge Impulse’s EON 120 ms 153 KiB
MIT’s TinyEngine 1 124 ms 98 KiB
STMicroelectronics’ X-CUBE-AI 103 ms 109 KiB
Plumerai’s inference software 77 ms 80 KiB

Model: MobileNetV2 2 3 (alpha=0.30, resolution=80x80, classes=1000)
Board: STM32F746G-Discovery at 216 MHz with 320 KiB RAM and 1 MiB flash

Our inference software builds upon TensorFlow Lite for Microcontrollers, such that it supports all of the same operations and more. But since resources are scarce on a microcontroller, we do not rely on TensorFlow or Arm’s kernels for the most performance-critical layers. Instead, for those layer types we developed custom kernel code, optimized for lowest latency and memory usage. This includes optimized code for regular convolutions, depthwise convolutions, fully-connected layers, various pooling layers and more. To become faster than the already heavily optimized Arm Cortex-M specific CMSIS-NN kernels, we had to go deep inside the inner-loops and also rethink the higher-level algorithms. This includes optimizations such as hand-written assembly blocks, improved register usage, pre-processing of weights and input activations and template-based loop-unrolling.

Although these generic per-layer-type optimizations resulted in great speed-ups, we went further and squeezed out every last bit of performance from the Arm Cortex-M microcontroller. To do that, we perform specific optimizations for each layer in a neural network. For instance, rather than only optimizing convolutions in general, our inference software makes specific improvements based on all actual values of layer parameters such as kernel sizes, strides, padding, etc. Since we do not know upfront which neural networks our inference software might run, we make these optimizations together with the compiler. This is achieved by generating code in an automated pre-processing step using the neural network as input. We then guide the compiler to do all the necessary constant propagation, function inlining and loop unrolling to achieve the lowest possible latency.

Memory usage is an important constraint on embedded devices; however fast or slow the software is, it has to fit in memory to run at all. TensorFlow Lite for Microcontrollers already comes with a memory planner that ensures a tensor only takes up space while there is a layer using it. We further optimized memory usage with a smart offline memory planner that analyzes the memory access patterns of each layer of the network. Depending on properties such as filter size, the memory planner allows the input and output of a layer to partially or even completely overlap, effectively computing the layer in-place.

Besides Arm Cortex-M, we also optimize our inference software for Arm Cortex-A and RISC-V architectures. And if the above results are still not fast enough for your application, we go even further. We make our AI tiny and radically more efficient by using Binarized Neural Networks (BNNs) - deep learning models that use only a single bit to encode each weight and activation. We are building improved deep learning model architectures and training algorithms for BNNs, we are designing a custom IP-core for customers with FPGAs and we are composing optimized training datasets. All these improvements mean that we can process more frames per second, save more energy, run larger and more accurate AI models and deploy on cheaper hardware.

Get in touch if you want to use the world’s fastest inference software and the most advanced AI on your embedded device.

Wednesday, October 13, 2021: Updated memory usage to match newest version of the inference software.


  1. Results copied from https://github.com/mit-han-lab/tinyml/tree/master/mcunet, all other results were measured by us. ↩︎

  2. The explicit padding layers in MobileNetV2 were fused, and to be able to compare with the TinyEngine results the number of filters in the final convolution layer (1280) was scaled by alpha to 384. The exact model can be downloaded here. ↩︎

  3. microTVM ran out of memory, other benchmarks show that microTVM is generally a bit slower than CMSIS-NN. ↩︎