Eta Compute Inc., a company dedicated to delivering machine learning to mobile and edge devices using its revolutionary new platform, today announced the availability of its latest machine learning SoC that includes autonomous learning. Named TENSAI, the product performs image classification, keyword spotting, and wakeup word detection that redefines the standard for ultra-low power embedded solutions.
The TENSAI chip includes the third generation of Eta Compute’s delay insensitive logic which enables products to reliably operate at the lowest supply voltage resulting in the lowest power consumption.
Other unique features of this SoC include:
- Eta Compute’s own kernel for spiking neural network (SNN) and CNN minimizes operations and lowers 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
- Image classification application consuming only 0.4mJ per picture, a 30X power reduction over recently published results
- Always-on wake-up word application which consumes 500uA during classification or 50uA during silence meeting strict requirements for wearables and battery-operated consumer electronics
“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.”
The processor is trainable using the popular TensorFlow or Caffe software and Eta Compute’s custom kernel further optimizes the trained model. TENSAI uses a tightly integrated DSP processor and microcontroller architecture for a significant reduction in power for embedded machine intelligence. This solution can support a wide range of applications in audio, video, and signal processing where power is a severe constraint as in mobile devices, wearable, industrial sensing, and camera markets.
Furthermore, for real-world scenarios for which readily labeled data is scarce or unavailable, our autonomous learning algorithms can extract actionable intelligence despite this limitation. This makes Eta Compute’s solution much broader in scope including intelligence for devices that harvest energy in remote environments.
Eta Compute SoC with machine learning is sampling now with mass production expected in Q1 of 2019.
Leave a Reply
You must be logged in to post a comment.