aiWare’s power efficient architecture enables neural network acceleration at the edge. The hardware IP core can be implemented in a wide range of smart sensors and sensor subsystems delivering 0.25–1 TMAC/s. In similar implementation the aiWare core can be expected to consume <1 W. Its low latency and deterministic behavior enable hardware platforms prepared for the challenges of production autonomous vehicles providing high-performance low-power, small size ECUs. Due to the wide range of ASIL-B–D compliant implementation options available the aiWare hardware IP core can be utilized in a wide range of automated driving solutions.
Due to its highly scalable architecture aiWare can be deployed as a gateway neural network accelerator as a part of automated driving systems. Relying on the aiWare core to pre-processes sensor data for sensor fusion maximises offload in the central processing cluster to enhance the robustness of ADAS functionalities. Offering a performance of 5–30 TMAC/s, such aiWare-based system can be tasked with perception and basic decision making tasks. Due to its highly optimized nature the power draw of the aiWare core in these implementations is 5–10 W. As aiWare contains no caches or programmable engines the IP core delivers continuous deterministic behavior.
The aiRing systolic ring structure available for high-performance aiWare implementations supports scaling up to 50 TMAC/s without compromising aiWare’s inherent characteristics. aiWare is capable of delivering a DL performance comparable to that of ten 30 TOP/s GPUs with a power consumption of less than 20 W. As aiWare is constantly tested against the aiDrive self-driving stack developed by AImotive the hardware IP core is benchmarked against real world autonomous driving workloads. As a result, aiWare is uniquely optimized to perform as an element of a self-driving compute platform.