The CNN engine delivers up to 4.5 TeraMACs per second when implemented in 16nm FinFET process technologies under typical conditions, four times more performance than the company’ previous CNN engine.
It also supports both coefficient and feature map compression/decompression to reduce data bandwidth requirements and decrease power consumption. The vision CPU scales from one to four vector DSPs and operates in parallel to the CNN engine, delivering maximum throughput for a broad range of high-performance embedded vision applications such as advanced driver assistance systems (ADAS), video surveillance, augmented and virtual reality, and simultaneous localization and mapping (SLAM).
The DesignWare EV6x Processor family integrates scalar, vector DSP and CNN processing units for highly accurate and fast vision processing. The EV6x supports any convolutional neural network, including popular networks such as AlexNet, VGG16, GoogLeNet, Yolo, Faster R-CNN, SqueezeNet and ResNet. Designers can run CNN graphs originally trained for 32-bit floating point hardware on the EV6x’s 12-bit CNN engine, significantly reducing the power and area of their designs while maintaining the same levels of detection accuracy. The engine delivers power efficiency of up to 2,000 GMACs/sec/W when implemented in 16-nm FinFET process technologies (worst-case conditions). The EV6x’s CNN hardware also supports neural networks trained for 8-bit precision to take advantage of the lower memory bandwidth and power requirements of these graph types.
To simplify software application development, the EV6x processors are supported by a comprehensive suite of tools and software. The latest release of the DesignWare ARC MetaWare EV Development Toolkit includes a CNN mapping tool that analyzes neural networks trained using popular frameworks like Caffe and Tensorflow, and automatically generates the executable for the programmable CNN engine. For maximum flexibility and future-proofing, the tool can also distribute computations between the vision CPU and CNN resources to support new and emerging neural network algorithms as well as customer-specific CNN layers. Combined with software development tools based on OpenVX, OpenCV and OpenCL C embedded