Eta Compute exhibiting at Arm TechCon, releases TENSAI – Machine Learning SoC ultra-low-power consumption solution with autonomous learning feature
Eta Compute Inc., major company providing machine learning (ML) solutions to mobile and edge devices, announces ML SoC named TENSAI, based on autonomous learning. This new product is able to do image classification, keyword spotting and wakeup word detection and all these features has revamped the set standard for ultra-low-power embedded devices.
The ML SoC is abased on third generation delay insensitive logic of Eta Compute, which allows the products to operate at the lowest voltage leading to lowest power consumption in a reliable way.
Important features of TENSAI:
- Spiking Neural Network (SNN) and CNN are Eta Compute’s own kernel which help in minimizing operations and reducing the power consumption
- Autonomous Learning of speech, image, and other data where classification occurs on the data without labels enabling advances in the broad area of anomaly detection on systems where failure modes are unknown or data difficult to obtain
- 0.4mJ consumption per picture for image classification and 30X power reduction over latest published results
- Always-on on wakeup word detection and this application consumes 500uA during classification or 50uA is required during silence meeting strict requirements for wearables and battery-operated electronic devices.
By using TensorFlow or Caffe software processor can be trained and custom kernel of Eta Compute further optimizes the trained model. A compact structure of DSP processor and microcontroller architecture is used in TENSAI chip which results into significant power reduction for embedded machine intelligence. A wide range of applications in audio, video and signal processing can be supported by this solution where power is a major constraint such as mobile devices, wearable, industrial sensing, and camera markets.
In the real world situation there is scarcity of readily labelled data or mostly its is unavailable, but despite of this limitation, the autonomous learning algorithms can extract actionable intelligence. This is a broader scope through which Eta Compute’s solution includes intelligence for devices that yields energy in remote environments.
Eta Compute’s TENSAI is right now in sampling stage and in the first quarter of 2019 it will start its mass production.
“Our patented hardware architecture (DIAL™) is combined with our fully customizable algorithms based on both CNN and SNNs to perform machine learning inferencing in hundreds of microwatts,” said Nara Srinivasa Ph.D., CTO of Eta Compute. “These are being sampled to customers who are integrating them into products such as smart speakers and object detection platforms to deliver machine intelligence to the network edge.”
“I know machine learning on tiny, cheap battery powered chips is coming,” said Pete Warden, Google Technical Lead of TensorFlow. This will open the door for some amazing new applications.”