TENSAI Flow enables development of machine learning

“Neural network and embedded software designers are seeking practical ways to make developing machine learning for edge applications less frustrating and time-consuming,” said Ted Tewksbury, CEO of Eta Compute. “With TENSAI Flow, Eta Compute addresses every aspect of designing and building a machine learning application for IoT and low power edge devices.

“Now, designers can optimise neural networks by reducing memory size, the number of operations, and power consumption, and embedded software designers can reduce the complexities of adding AI to embedded edge devices, saving months of development time.”

Eta Compute’s TENSAI Flow software de-risks development by quickly confirming feasibility and proof of concept. TENSAI Flow enables development for machine learning applications in IoT and low power edge devices. It includes a neural network compiler, a neural network zoo, and middleware comprising FreeRTOS, HAL and frameworks for sensors, as well as IoT/cloud enablement. 

“In order to best unlock the benefits of TinyML we need highly optimised hardware and algorithms. Eta Compute’s TENSAI provides an ideal combination of highly efficient ML hardware, coupled with an optimised neural network compiler,” said Zach Shelby, CEO of Edge Impulse. “Together with Edge Impulse and the TENSAI Sensor Board this is the best possible solution to achieve extremely low-power ML applications.”

The TENSAI Flow exclusive neural network compiler delivers the best optimisation for neural networks running on Eta Compute’s device as well as power efficiency. In addition, the middleware makes dual core programming by eliminating the need to write customised code to take full advantage of DSPs.

A Neural Network Zoo accelerates and simplifies development with ready-to-use networks for the most common use cases. These will include motion, image and sound classification. Developers simply train the networks with their data. And, with the insight from TENSAI Flow’s real world applications, developers can see the potential of neural sensor processors precisely in terms of energy efficiency and performance in a variety of field tested examples with unmatched efficiency while preserving total flexibility.

Compared to direct implementation on a competitive device of the same CIFAR10 neural network, the TENSAI neural network compiler on TENSAI SoC improves energy per inference by a factor 54x. Using the CIFAR10 neural network from TENSAI neural network zoo and TENSAI neural network compiler improves the energy per inference further, bringing it to a 200x factor.

"Google and the TensorFlow team have been dedicated in bringing machine learning with the tiniest devices. Eta Compute's TENSAI Flow is another step in the same direction and enables TensorFlow networks to run on Eta Compute's ultra low power SoC, with the best optimisation the company  can provide,” said Pete Warden, Lead of the TensorFlow Mobile/Embedded team at Google. “We welcome this initiative that sets new benchmarks for machine learning in edge devices and shows the dynamism of the TinyML field."

Through its interface with Edge Impulse, TENSAI Flow allows developers to securely acquire and store training data so customers train once and have real-world models for future development. The software automatically optimises TensorFlow Lite AI models for Eta Compute’s TENSAI SoC, delivering high optimisation and power efficiency. With TENSAI Flow, TENSAI SoC can load AI models that include sensor interfaces. TENSAI Flow provides the foundation to automatically provisions and connects devices to the cloud and upgrades firmware over the air based on new models or data.

“The TENSAI Platform enables AI performance in the range of one milliwatt, which is extremely low compared to other solutions, especially for image processing,” said Jim Feldhan, President and Founder of Semico Research. “Eta Compute’s technology eliminates power consumption as a barrier for AI applications at the Edge.”