CEVA, the world's leading signal processing IP licensing company focused on smart connected devices, announces a new DSP-based product that brings deep learning and artificial intelligence (AI) capabilities to low-power embedded systems. At the heart of this comprehensive, scalable, integrated hardware and software IP platform is the new image and vision DSP CEVA-XM6, enabling developers to efficiently utilize neural network and machine vision capabilities for smartphones, unmanned vehicles, surveillance, Robots, drones and other smart devices with camera functions.
Compared to the previous generation CEVA-XM4 intelligent vision DSP, the CEVA-XM6-based new vision platform has eight times better neural network performance and three times the performance improvement of all computer vision cores. Key enhancements to this new architecture integration include new vector and scalar processing units, as well as extensive enhancements to the instruction set, memory bandwidth, and DMA.
This new vision platform further extends the performance benefits of CEVA products when implementing neural networks compared to mainstream architectures using GPUs. Compared to leading GPU embedded systems for computer vision and deep learning, CEVA's latest image and vision platforms deliver up to 25 times better performance-per-watt efficiency for convolutions such as AlexNet and GoogLeNet. The neural network (CNN) is four times faster.
"With computer vision and deep learning technology becoming mainstream, we need deep neural networks generated by high-power GPU engines and deploy these embedded applications with limited power and performance," said Ilan Yona, vice president and general manager of CEVA's vision business unit. Building bridges between them to eliminate the gap in the middle. Our new vision platform is excellent in this regard, providing developers with the most comprehensive set of technologies to quickly respond to these embedded use cases."
This vision platform integrates a wide range of software and hardware IP to provide time-to-market and power advantages for deploying machine vision and deep learning in embedded systems. In addition to the CEVA-XM6 DSP itself, the platform includes CNN-specific feature accelerators and image de-warping (for all types of image transformation), CEVA's widely acclaimed CDNN2 neural network software framework, OpenCV, OpenCL and OpenVX API, CEVA- CV computer vision library, and a set of widely used optimization algorithms.
Jeff Bier, founder of the Embedded Vision Alliance, said: "Designers of various end products are eager to integrate visual intelligence into their designs. Typically, the visual and deep learning algorithms used by these developers require low cost, low power consumption, and The programmable approach provides extremely high processing performance, and I appreciate CEVA's long-standing commitment to providing processors and software tools that meet these needs."
Technical characteristicsThe CEVA-XM6 is based on the powerful CEVA-XM4 and CEVA-MM3101 processors and has more than 25 design projects. It features a range of architectural innovations and enhancements that provide breakthrough neural network performance and advanced computer vision processing capabilities, including:
Innovative Vector Processing Unit (VPU) architecture: ensuring more than 95% MAC utilization is an unparalleled level of excellence in the industry today
· Enhanced Parallel Decentralization – Centralized Storage Load Mechanism: Further improve visual algorithm performance, including SLAM and depth mapping.
Sliding Window 2.0TM: This patented mechanism leverages pixel overlays in image processing to help achieve higher utilization in a wider range of neural networks and to accommodate the increasing complexity of these networks.
Option 32-channel SIMD vector floating-point unit, including IEEE semi-precision standard (FP16) and significant nonlinear computation enhancements.
· Other improvements include an enhanced 3D data processing solution that accelerates CNN performance, a 50% improvement in control code performance compared to CEVA-XM4, and new scalable units and multi-core and system integration support to further reduce code size.
In addition to the CEVA-XM6 DSP, other key components of this vision platform include:
· CDNN Accelerator: The 16-bit CDNN Accelerator has 512 MACs/cycle, ensuring the industry's best performance to handle the most complex neural networks available today. The CDNN Accelerator is also used to release 256 MAC units within the CEVA-XM6 DSP, allowing other computer vision tasks to run in parallel. This flexible approach makes the CDNN Accelerator with the CEVA-XM6 architecture the best choice for supporting new image algorithms, network architectures, and changing the layer types in the rapidly evolving deep learning space.
· Image De-warping Accelerator: For wide-angle camera applications, such as 360-degree cameras, the Image De-warping Accelerator supports the ARM Frame Buffer Compression (AFBC) protocol for optimal system interoperability.
Accelerator Cognitive Supplement: Run on the CEVA-XM6 DSP to provide efficient accelerator utilization, enabling designers to further differentiate their product designs.
· CDNN2 software framework: Optimized and collaborated with CEVA-XM6 and accelerators, making it easy for developers to use this tool to generate and port their proprietary neural networks to CEVA-XM6, significantly accelerating the use of the latest and most advanced network topologies And layer performance, including support for Caffe and Google's machine learning software library TensorFlow.
· Proof of ISO 26262 active safety: the need to support next-generation ADAS and autonomous driving solutions in automotive applications.
CEVA's CDNN2 software framework is optimized for both CEVA-XM6 and CDNN accelerators, fully supporting 16-bit fixed-point accuracy, ensuring less than 1% accuracy when running networks trained in 32-bit floating-point environments. This is the key to the transition from neural networks to high-cost and power-efficient solutions for high-volume automotive and consumer applications.
SupplyCEVA will license CEVA-XM6 DSP and vision platform components to major customers in the fourth quarter of 2016 and will be licensed in the first quarter of 2017.
To learn more about the CEVA-XM6 DSP and vision platform, visit the company website. A video introducing CEVA-XM6 can be viewed on the CEVA YouKu channel.
64V Battery Pack ,Lithium Battery Box,Lithium Power Pack,Jackery Battery Pack
Zhejiang Casnovo Materials Co., Ltd. , https://www.casnovonewenergy.com