Arm NPUs deliver high throughput and efficiency in AI applications on the edge. They provide significant advantage in power, performance, and area over conventional CPUs and GPUs for INT8 inferences out-of-the-box. However, further performance improvements can be unlocked by fine-tuning neural network models to take advantage of the underlying hardware architecture. This presentation will describe the recent developments in open-source machine learning tools that enable such optimizations and explore how various techniques can be combined to achieve the optimal balance between model accuracy and inference performance.

In this presentation we will discuss the importance of software in maximizing the throughput, accuracy, and power of an AI inference accelerator. We examine how codeveloping software with hardware allows for architecture tradeoffs that maximize throughput/power for customer models. The software compiler must seamlessly translate data into meaningful results without knowledge of the hardware inner workings. Finally, with the continued evolution of CNN models, software adaptability will continue to drive throughput/power/cost improvements for broader adoption of AI functionality.

In order to accelerate the execution of a neural network using an AI processor, a customer must first be able to get their neural network up and running. The software part of the solution is a critical, and often under-estimated, factor in the semiconductor world. This presentation will focus on the Mythic Optimization Suite and Graph Compiler, the tool flow developed by Mythic to bring up neural networks on the innovative Mythic Analog Matrix Processor (AMP) architecture, as well as the unique features and capabilities of the Mythic AMP itself.

The demand for ultra-low power and incremental learning is key to the future of AI.  Advanced neuromorphic computing is delivering a pathway to new technologies that are driving the ecosystem. Solving problems in machine learning such as privacy, latency, and power, neuromorphic computing is challenging the current way of AI thinking. This presentation addresses the ease of converting CNNs to the event domain by processing in the MetaTF neuromorphic environment from BrainChip, enabling designers to achieve efficient and effective AI solutions.

Advanced safety features for ADAS and autonomous driving capabilities are two automotive megatrends. Radar is a key component to implementing such systems. This talk will review the attributes of key silicon technologies for mmWave radar applications. Specific design examples will be given that highlight the performance and integration capabilities of the technologies that meet the stringent requirements of automotive OEMs.

Wireless audio SoC design requires rare multidisciplinary technology expertise. CEVA addresses this need by offering a solution that dramatically lowers the high entry barriers to the True Wireless Stereo (TWS) audio market. This talk will present Bluebud, a turnkey IP platform based on the RivieraWaves Bluetooth 5.2 IP, system peripherals, and a single CEVA-BX1 DSP core running a complete software framework comprising a Bluetooth stack along with an advanced stereo synchronization mechanism, audio codecs, and a selection of voice and motion processing software add-ons.