Thanks AD. I know it is an old 2018 article but as you know I embrace the old. LOL
My opinion only DYOR
FF
AKIDA Ballista
AI in embedded use: first neuromorphic SoC component announced
A new type of acceleration SoC is intended to bring neural networks and artificial intelligence (AI) to the edge of networks and in companies: With the Aikida, Brainchip has announced the first SoC available in series, specially designed for SNNs (Spiked Neural Networks).
Companies on the topic
BrainChip introduces the Akida architecture, a neuromorphic system-on-chip component designed for so-called Spiked Neural Networks (SNN).
(Image: Brainchip)
BrainChip is the first company that wants to bring a pulsed neural network architecture (SNN; Spiking Neural Network) to the market in series: the neuromorphic system-on-chip (NSoC) Akida.
"The market for AI accelerator ICs will exceed US $ 60 billion by 2025," states Aditya Kaul, Research Director at Tractica, a market research company specializing in artificial intelligence.“Neuromorphic computers promise faster AI, especially in low-power applications.Since many technical hurdles have now been solved, the industry will use a new class of AI-optimized hardware in the next few years. "
"Despite great efforts, no other company has succeeded in bringing a neuromorphic IC to the market in series," said Lou DiNardo, CEO of BrainChip.“Akida, Greek for 'pulse / peak', is the first component of a new generation of AI hardware solutions.AI at the edge of the network will be as important and useful as the microcontroller. "
Emulation of neurons
The Akida NSoC relies on so-called pulsed neural networks (Spiked Neural Networks) instead of the convolutional neural networks (CNN) that are currently more common."SNNs are considered to be the third generation of neural networks," says 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."
The idea: SNNs should emulate neurons in the human brain that transmit data in the form of impulses.This ensures a more efficient, less energy-hungry way of working.Instead of continuously transmitting the entire input of all sensor data, for example, an impulse is only sent when something changes."Changes are events," explains van der Made. "Whenever an event occurs - for example a change in a pixel detected by a sensor - an impulse is generated and sent on along the neuron path."So only this event, this pixel change is transmitted - not the overall condition of the image.
Each Akida NSoC has an effective 1.2 million neurons and 10 billion synapses.This would promise 100 times better efficiency than, for example, theLoihi test chip from Intel presented last year.Comparisons with leading CNN accelerators show increases in performance of more than an order of magnitude in image / seconds / watt benchmarks such as CIFAR-10 with comparable accuracy: While Loihi would only achieve below 20 fps per watt, Akida would already have 1400 fps / W.
Standalone for embedded or co-processor applications
The Akida NSoC was developed for use as a stand-alone embedded accelerator or as a co-processor.It contains sensor interfaces for conventional pixel-based imaging, dynamic image sensors (DVS), lidar, audio and analog signals.It also has high-speed data interfaces such as PCI-Express, USB and Ethernet.The NSoC also contains data-to-pulse converters that convert common data formats optimally into pulses / spikes in order to be trained and processed by the Akida neuron fabric.
The Akida NSoC is designed to enable off-chip training - or on-chip training - in the Akida development environment.An integrated ARM core processor controls the configuration of the Akida neuron fabric as well as the off-chip communication of metadata.The Akida Development Environment is available immediately for Early Access customers who want to begin creating, training, and testing Akida NSoC-based SNNs.The Akida NSoC is expected to be available as a sample in the third quarter of 2019.The targeted price for series production should be US $ 10 each.
(20min delay)