Akida should sample in fall 2019, delivering nearly an order of magnitude more throughput/watt than a Movidius Myriad 2 at about the same price and accuracy.
SNNs are a form of machine learning, related but different from the convolutional neural nets now widely used by web giants for jobs like voice and image recognition. SNNs use a simpler, one-shot training method and are well-suited to tasks such as face recognition in low-resolution and noisy environments such as surveillance video.
The Akida NSoC is small, low cost and low power, making it a good solution for edge applications such as advanced driver assistance systems (ADAS), autonomous vehicles, drones, vision-guided robotics, surveillance and machine vision systems. Its scalability allows users to network many Akida devices together to perform complex neural network training and inferencing for many markets including agricultural technology (AgTech), cybersecurity and financial technology (FinTech).
The Akida NSoC uses a pure CMOS logic process. Spiking neural networks (SNNs) are inherently lower power than traditional convolutional neural networks (CNNs), as they replace the math-intensive convolutions and back-propagation training methods with biologically inspired neuron functions and feed-forward training methodologies. BrainChip's research has determined the optimal neuron model and training methods, bringing efficiency and accuracy. Each Akida NSoC has effectively 1.2 million neurons and 10 billion synapses, representing 100 times better efficiency than neuromorphic test chips from Intel and IBM. Comparisons to leading CNN accelerator devices show similar performance gains of an order of magnitude better images/second/watt running industry standard benchmarks such as CIFAR-10 with comparable accuracy.
"Spiking neural networks are considered the third generation of neural networks," said Peter van der Made, Founder and CTO of BrainChip. "The Akida NSoC is the culmination of decades of research to determine the optimum neuron model and innovative training methodologies."
The Akida NSoC is designed for use as a stand-alone embedded accelerator or as a co-processor. It includes sensor interfaces for traditional pixel-based imaging, dynamic vision sensors (DVS), Lidar, audio, and analog signals. It also has high-speed data interfaces such as PCI-Express, USB, and Ethernet. Embedded in the NSoC are data-to-spike converters designed to optimally convert popular data formats into spikes to train and be processed by the Akida Neuron Fabric.
Spiking neural networks are inherently feed-forward dataflows, for both training and inference. Ingrained within the Akida neuron model are training methodologies for supervised and unsupervised training. In the supervised mode, the initial layers of the network train themselves autonomously, while in the final fully-connected layers, labels can be applied, enabling these networks to function as classification networks. The Akida NSoC is designed to allow off-chip training in the Akida Development Environment, or on-chip training. An on-chip CPU is used to control the configuration of the Akida Neuron Fabric as well as off-chip communication of metadata.
The Akida Development Environment is available now for early access.
BrainChip got its start 10 years ago as a spinout from the University of Toulouse research effort that was creating custom software for users in France. BrainChip now consists of a software team in Toulouse, a hardware group in southern California, and last year, it brought on new management mainly in Silicon Valley.