In the data center, new architectures are emerging to challenge the GPU's dominance in AI training and inference. Embedded systems such as smart cameras, smart vehicles, and smart robots require powerful accelerators. AI accelerators are even moving into IoT and smart-home devices, running simple neural networks on milliwatts of power. This presentation will describe the latest trends in AI acceleration while addressing how these accelerators are used across this range of end applications.

This presentation will introduce product details of the ADK1000, BrainChip's first event domain neural-network IP and SoC (NSoC) device, which enables AI capability in edge IoT systems at ultra-low power consumption. Based upon the fundamentals of neuromorphic computing, the neural processor can run a standard CNN by converting it into the event-domain, allowing it to implement transfer learning and incremental learning on the chip. The same neural processor can also train and run SNNs natively.

Edge machine learning generally involves real-time video or audio feeds that exhibit great sparsity – sparsity in time, as information rarely changes; sparsity in space, as large sections of visual or audio space contain no information; and sparsity in processing, as edge neural networks can have a majority of inactive connections in any processing period. This presentation will show how GrAI Matter Labs' architecture can exploit all three types of sparsity to achieve new levels of efficiency in edge processing.

The intersection of IoT, AI and 5G is driving the need for more on-device intelligence in smaller, cost-sensitive devices. Arm recently announced new ML IP for microcontrollers: the Cortex-M55 processor, the first to feature Helium technology, and the Arm Ethos-U55 microNPU (neural processing unit), the industry's first microNPU designed to accelerate ML performance. This presentation will describe the processors' features, configuration options, system design considerations, and how they can be used to address a wide range of use cases.