Bookmark and Share

Cadence Demonstration of a Neural Network on the Vision P6 DSP

Megha Dagha , Senior Technical Marketing Manager at Cadence, demonstrates the company's latest embedded vision technologies and products at the May 2017 Embedded Vision Summit. Specifically, Daga demonstrates Cadence’s highly programmable, efficient Tensilica® Vision P6 DSP by showing an implementation of one of the most common neural networks, AlexNet. The demo uses a trained floating-point neural network from Caffe. Tensilica provided the most optimal solution by using their custom quantization to 8 bits with less than 1% accuracy drop. By using quantized coefficients, the Vision P6 DSP was able to achieve low bandwidth, low power, and high performance at minimal degradation.