Hello everyone, I am a ruffian balance, is a serious technical ruffian. Today ruffian balance to introduce you is ARM Ethos-U55.

ARM has just released the cortex-M family’s latest kernel, the Cortex-M55, and the first Micronpu-Ethos for the Cortex-M series. Cortex-m55 is the first cortex-M kernel for AI/ML, Rascal Heng has also written a short article for this purpose “born for AI/ML (Cortex-M55)”, in that article Rascal Heng is just a brush with Ethos-U55, did not discuss in depth, Today rascal balance with you to have a good chat about Ethos-U55.

Before introducing Ethos-U55, it is worth mentioning ARM’s Project Trillium, a machine learning platform that ARM launched in February 2018, including a new ARM IP portfolio of highly scalable processors and associated software. The platform includes the Mali SERIES GPU and Arm NN libraries, as well as the Ethos-N SERIES NPU, which was released in October 2019. Now, Ethos-U55 has officially joined Project Trillium.

So what exactly is Ethos-U55? It is actually a small, dedicated microNPU inference accelerator, with powerful computing power at its core, up to 64-512 GOP/s (Giga Operations Per Second), This computing power beats the Cortex-M CPU. But its Trillion Operations Per Second (Trillion Operations Per Second) is still weaker than the large NPU-Ethos-N used on the Cortex-A series, so it’s microNPU, Works well with the Cortex-M.

On the neural network framework, Ethos-U55 currently only supports TensorFlow Lite Micro. This is also a bit boring compared with the Framework family bucket (TensorFlow, TensorFlow Lite, Caffe2, PyTorch, MXNet, ONNX) supported by Ethos-N. However, it is the SINGLE-chip AI after all. The first step is to be able to play. TensorFlow Lite is enough.

Although the Ethos-U55 is released with the Cortex-M55, it doesn’t just work with the Cortex-M55. It also works with the Cortex-M4/M7/M33. However, the Ethos-U55 works best with the Cortex-M55 system, with combined performance up to 50 times better than the Cortex-M7 system alone, and ML performance up to an eye-popping 480 times better.

With all that talk about the Ethos-U55 features, let’s take a look at how it works with the Cortex-M, taking its best partner, the Cortex-M55, as an example. There are two AXI master interfaces (64bit) designed on Ethos-U55. Axi-1 is readable and writable for accessing the shared SRAM. In addition, AXI-2 is read-only and is specially used to fetch instructions from NVM(Flash). In addition, an APB Slave interface is designed on Ethos-U55 to provide access to control and configuration registers.

Therefore, the workflow of Ethos-U55 in the system is as follows: Cortex-M55 sends ethos-U55 startup neural network processing commands through APB interface, and Ethos-U55 obtains specific command sequences from NVM through AXI-2 after receiving the commands. After processing, Ethos-U55 sends an interrupt event to cortex-M55 through a dedicated IRQ signal cable.

Let’s take a typical intelligent voice application to analyze the cooperation mechanism between Ethos-U55 and Cortex-M55. By default, Ethos-U55 maintains low power consumption mode, and Cortex-M55 completes speech detection and wakeup word recognition (KWS). Once effective wakeup words are detected, At this point, Ethos-U55 can be activated to complete subsequent speech recognition (ASR) tasks and further neural network processing to identify semantics.

We don’t know which MCU manufacturer will be the first to launch an MCU with this microNPU. We don’t know which MCU with this microNPU will be paired with which Cortex-M kernel for the first time.

Welcome to subscribe to

The article will be published on my blog park homepage, CSDN homepage and wechat public account platform at the same time.

Wechat search “ruffian balance embedded” or scan the following two-dimensional code, you can see the first time on the phone oh.