Synopsys has developed a new family of embedded vision processor designs to support high performance neural networks, writes Nick Flaherty.
The EV52 and EV54 vision processors are fully programmable and configurable vision processor IP cores that combine the flexibility of software solutions with the low cost and low power consumption of dedicated hardware.
The EV Processors implement a convolutional neural network (CNN) that can operate at more than 1000 GOPS/W, enabling fast and accurate detection of a wide range of objects such as faces, pedestrians and hand gestures at a fraction of the power consumption of competing vision solutions.
To speed application software development, the EV Processor Family is supported by a comprehensive software programming environment based on existing and emerging embedded vision standards including OpenCV and OpenVX, as well as Synopsys’ MetaWare Development Toolkit. The combination of high-performance hardware optimized for vision data processing and high productivity programming tools make the EV Processors a solution for a broad range of embedded vision applications including video surveillance, gesture recognition and object detection.
“The ability of embedded systems to extract meaning from visual inputs is becoming increasingly important in a broad range of products including security equipment, gaming devices and automobiles. This is driving demand for more performance- and power-efficient vision processing capabilities,” said Jeff Bier, founder of the Embedded Vision Alliance. “Specialized processors like the Synopsys’ DesignWare EV Processors help designers attain the performance they need for their vision applications at levels of power consumption suitable for portable devices.”
The EV Processors include multiple high-performance processing cores that can operate at up to 1 GHz in typical 28-nanometer process technologies. The EV Processors also implement a feed-forward CNN structure that supports a programmable point-to-point streaming interconnect for fast and accurate object detection, a critical task in vision processing.
The processors’ configurable number of execution units enable developers to exploit the task- and data-level parallelism common in vision applications, executing complex image and video recognition algorithms with as little as one-fifth the power consumption of other vision processors available on the market.
A complete software programming environment, including OpenVX and OpenCV libraries, and Synopsys’ MetaWare Development Toolkit, simplifies the development of application software for the Synopsys EV Processor Family. The OpenCV source libraries available for EV Processors provide more than 2500 functions for real-time computer vision.
The processors are programmable and can be trained to support any object detection graph. The OpenVX framework includes 43 standard computer vision kernels that have been optimised to run on the EV Processors, such as edge detection, image pyramid creation and optical flow estimation.
Users can also define new OpenVX kernels, giving them flexibility for their current vision applications and the ability to address future object detection requirements. The OpenVX runtime distributes tiled kernel execution over the EV Processors’ multiple execution units, simplifying the programming of the processor.
The full suite of tools and libraries, along with available reference designs, enable designers to efficiently build, debug, profile and optimise their embedded vision systems.
The EV Processors are designed to integrate seamlessly into an SoC. They can be used with any host processors and operate in parallel with the host. The EV Family includes support for synchronisation with the host through message passing and interrupts.
In addition, the EV Processor memory map is accessible to the host. These features enable the host to maintain control while allowing all vision processing to be offloaded to the EV Processor, reducing power and accelerating results.
The EV Processors can access image data stored in a memory mapped area of the SoC or from off-chip sources independently from the host through the ARM AMBA AXI standard system interface if required.
“Embedded vision is driving innovation in a broad spectrum of applications, from surveillance to consumer and gaming devices,” said John Koeter, vice president of marketing for IP and prototyping at Synopsys. “The EV Processor Family delivers state-of-the-art object detection accuracy with 5X better power efficiency, along with comprehensive vision libraries and a robust software programming environment. This combination enables design teams to integrate embedded vision functionality into more systems faster with much lower power consumption than existing solutions.”