Eta Compute’s tool chain is growing, with the announcement of a compiler for its ECM3532 chip that streamlines embedded development.
Combined with a sensor board evaluation kit the size of a coin, and a partnership with Edge Impulse, this pushes the startup to the next level of maturity. Eta Compute launched its ECM3532 ultra-low-power AI chip in February, but until now was supporting customer designs by manually optimizing code for the dual-core MCU and DSP design.
The ECM3532 is a dual-core SoC that uses Arm Cortex-M3 and NXP CoolFlux DSP cores for AI processing. The company uses a patented continuous voltage frequency scaling (CVFS) technique to adjust the voltage and clock frequency of both cores to meet the variable needs of IoT devices. It’s intended for sensor fusion applications in battery-powered designs; always-on image processing applications can be achieved with a power budget of 100 µW.
A tiny evaluation board launched recently is intended to ease the development of smart sensors, embedding microphones, temperature and pressure sensor, accelerometer, gyroscope and Bluetooth connectivity. It measures 1.4 by 1.4 inches and can run for “months” on a coin cell battery, according to the company.
Eta Compute’s Tensai Sensor Board a complete AI-enabled sensor node (Image: Eta Compute)
Eta Compute partnered with Edge Impulse back in May, with Eta Compute’s chip and evaluation board supported by Edge Impulse’s end-to-end ML development and MLOps platform. Much of Edge Impulse’s tooling handles visualization and management of datasets for AI-enabled IoT nodes.
“It can be really hard to start using [an Eta Compute] kind of part, as an embedded developer… and doing ML is even harder,” Edge Impulse CEO Zach Shelby told EE Times in an earlier interview. “We try to take that pain away. We have a nice drag and drop binary that goes on [Eta Compute] boards. It starts collecting sensor data right away, into our system. And then when it’s time to deploy the ML algorithm, we have a deployment option that builds a library for the Eta Compute target that will run right on the device.”
Compiler and middleware
Edge AI developers today face several problems that Eta Compute’s Tensai Flow tool chain addresses, according to Semir Haddad, Senior Director Product Marketing at Eta Compute.
“The first one is how you interface with real sensors and capture that data to improve your network,” said Haddad. “The second thing is how you optimize the network for the hardware. Today you have neural network frameworks or tools that help you run on CPUs, but it’s not really optimized [for our hardware]. Then, you need to generate firmware that can be used in real embedded system development. The fourth problem is the complete edge to cloud solution including device provisioning and connection to the cloud. These are the four pain points we are addressing with Tensai Flow.”
Eta Compute’s Tensai Flow tool chain now includes a compiler that optimizes neural network code for the company’s chip (Image: Eta Compute)
Tensai Flow includes a compiler that takes a TensorFlow or ONNX model and compiles it to code that can be executed on the ECM3532 device. Middleware adds all the software needed to run a complete application, including a real time operating system (RTOS) and sensor drivers.
Tensai Flow also includes a ‘network zoo’ of pre-validated models for specific use cases that can be integrated into customer designs. Edge Impulse handles data ops – collection and management of data, versioning of the data, sharing datasets between developers, etc.
“What is unique with this solution compared to what we can find from other vendors… is how comprehensive it is, in terms of being able to generate optimized code that can be used in a real application,” said Haddad. “Between the neural network aspect and the firmware development aspect, this differentiates [Tensai Flow].”
Founded in 2015, Eta Compute has achieved production silicon with a relatively modest budget and head count. The company has raised $19 million to date and has 35 staff in the US and India.
>> This article was originally published on our sister site, EE Times.