The company has now added a software and development stack under the name Tensai with some hardware enhancements allowing the single chip to be used for image classification, keyword spotting, and wakeup word detection. The processor can be trained using the popular TensorFlow or Caffe software.
The SoC includes the Cortex-M3 operating at up to 100MHz clock frequency and the NXP CoolFlux 8/16 bit dual MAC DSP. Additonal hardware has been includes to move data quickly between the processor and the DSP, said Paul Washkewicz vice president of marketing and co-founder of Eta Compute.
Eta provides software kernels for convolutional neural networks on the Coolflux DSP which are scalable to other NNs and that will cut a further 30 percent power on our production tapeout with asynchronous technology later in Q4. Tensai also provides a kernel to support proprietary spiking neural network design and tool development.
Spiking neural networks provide the opportunity for unsupervised learning. One example of this is autonomous learning of speech, image and other data, where classification can occur on data without labelling. This enables a broad range of anomaly detection where failure modes are unknown or data is difficult to obtain.
Comparison of CIFAR10 benchmark on STMicroelectronics and Eta Compute platforms. Source: Eta Compute.
Eta Compute claims Tensai offers a 30x power reduction in a specific CNN-based image classification benchmark – CIFAR10 – against an STMicroelectronics Cortex-M7 microcontroller. Eta Compute achieved 0.4mJ per picture across 8 million operations.
Next: CTO speaks
“Our patented hardware architecture 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 CTO of Eta Compute, in a statement. “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.”
Eta Compute classifies its Tensai MCU as an “edge device” claiming it can support full applications in a single device and differentiates it from some competition that are designed to be part of multiple compment node solutions.
Classification of “edge” and “node” neural network processors. Source: Eta Compute.
Eta Compute SoC with machine learning is sampling now with mass production expected in 1Q19.
Eta raises funds for spiking neural networks
BrainChip launches spiking neural network SoC
Neuromorphic computing runs on asynchronous Cortex-M3
Self-timed logic is Eta Compute’s low-power secret