It is claimed that Ztachip, an open source RISC-V accelerator, performs 20 to 50 times faster than non-accelerated RISC-V implementations and outperforms RISC-V cores with vector extensions in vision and AI edge applications operating on low-end FPGA devices or custom ASIC (no numbers were provided here).
Ztachip, pronounced “zeta-chip,” can speed up common computer vision tasks like edge detection, optical flow, motion detection, colour conversion, as well as TensorFlow AI models without retraining. It is not tied to any specific architecture, but the example code includes a RISC-V core based on the VexRiscv implementation.
An OV7670 VGA camera module, a PMOD VGA module for connecting to a display, and an ArtyA7-100T FPGA board from Digilent were used to evaluate the open source AI accelerator. The sample may then be built using the Xilinx Vivado Webpack free version according to the instructions provided on Github, and it can then be flashed to the board using OpenOCD.
The Ztachip AI vision accelerator is shown in action in the video below, executing a multi-tasking demonstration with object detection, edge detection, Harris corner detector, and motion detection all active at once.
The creator, Vuong Nguyen, claims that his accelerator is more adaptable and supports a wider variety of AI workloads than other accelerators, which tend to accelerate only a limited number of applications, such as convolution neural networks (CNN) alone. The project is free to use, even for commercial purposes, as it has been provided under an MIT licence.