Hailo’s Hailo-15 AI Vision Processor is built around the Hailo Neural Network (NN) core that can deliver up to 20 TOPS of performance (see figure). It’s designed to handle video streams up to 4k60 in real-time with the help of a vision subsystem and image signal processor (ISP).
The system-on-chip (SoC) isn’t just an AI accelerator. The application subsystem includes four 1.3-GHz Arm Cortex-A53 cores. A 32-bit LPDDR 4/4X memory interface runs at 4,266 Mtransfers/s.
The DSP vision-processing subsystem consists of a 700-MHz Vector DSP that can deliver 256 MACs and 350 GOPS. The ISP can handle up to a 12-Mpixel resolution with RGGB and RCCB color-filter-array (CFA) and wide-dynamic-range (WRD) support. In addition, it supports high dynamic range (HDR) and noise-reduction algorithms. The video system works with dual MIPI CSI inputs and a 24-bit digital video port (DVP).
Also in the mix are x4 PCI Express Gen 3 interfaces, 1-Gb Ethernet with RMII/RGMI interfaces, and USB 3.1 Gen2 host and device interfaces. The TrustZone security system supports secure boot and secure debugging. It also comes with a hardware-accelerated crypto library.
The chip is housed in a 15- × 15-mm FCBGA package. Operating temperature ranges from −40 to 85°C. A camera kit single-board computer is available.
The chip supports Yocto-based Linux. The Hailo Software Suite, data-flow compiler, and HailoRT toolkit are available from Hailo with optional Docker support.