This is because the new neuromorphic systems on a chip (NSoC) from the Akida family work on the basis of the spiking neural network (SNN) architecture instead of the convolutional neural networks commonly used today. The SNN processors are 100 times more effective than the CNN types, they are small, cost-effective and energy-saving.

BrainChip sees itself as a pioneer in bringing artificial intelligence (AI) to edge devices. Examples include advanced driver assistance systems (ADAS), autonomous vehicles, drones, image-controlled robotics as well as monitoring and image processing systems. Several Akida NSoCs can be cascaded to train complex neural networks and perform inferencing in numerous applications such as agricultural technology (AgTech), cyber security and financial technology (FinTech).

A new provider in a large future market

"The market for AI accelerator ICs will exceed 60 billion dollars by 2025," says Aditya Kaul, Research Director at Tractica, a market research company specializing in artificial intelligence. "Because many technical hurdles have now been overcome, the industry will be able to deploy a new class of AI-optimized hardware in the coming years."

"Despite our best efforts, no other company has succeeded in bringing a neuromorphic IC to market in series production," explains Lou DiNardo, CEO of BrainChip. "AI at the network edge will be as significant and useful as the microcontroller."

Biologically inspired

SNNs do not have to solve complex mathematical equations to find the weights for the neural networks, as is common with CNNs. They also do not have to be trained for the respective tasks as time-consuming and performance-intensive as CNNs.

Each Akida NSoC effectively integrates 1.2 million neurons and 10 billion synapses, which is 100 times more efficient than neuromorphic test chips from Intel and IBM. Comparisons with leading CNN accelerators show performance gains of more than an order of magnitude in frame/second/watt benchmarks such as CIFAR-10 with comparable accuracy.

"SNNs are considered the third generation of neural networks," said Peter van der Made, founder and CTO of BrainChip. "The Akida NSoC is the result of decades of research to determine the optimal neuron model and innovative training methods."

Embedded and co-processor

BrainChip has developed the Akida NSoC for use as a stand-alone embedded accelerator and as a co-processor. It includes sensor interfaces for conventional pixel-based imaging, dynamic vision sensors (DVS), lidar, audio and analog signals. It also features high-speed data interfaces such as PCI Express, USB and Ethernet.

Innovative training methods

SNNs are feed-forward data flows from the ground up - for training and inferencing. The Akida neuron model covers innovative training methods for supervised and unsupervised training. In supervised mode, the first layers of the network train themselves autonomously, while labels can be applied in the last fully connected layers. These networks thus serve as classification networks. The Akida NSoC is designed to enable off-chip training - and on-chip training - in the Akida development environment. An integrated CPU controls the configuration of the Akida neuron fabric as well as the off-chip communication of metadata.

The Akida development environment is available now for early access customers who want to start building, training and testing Akida NSoC-based SNNs. The Akida NSoC is expected to be available as a sample in the third quarter of 2019.