Fifth-generation imaging and vision technology simplifies deep learning
Targeted for autonomous driving, sense and avoid drones, virtual and augmented reality, smart surveillance, smartphones and robotics, a scalable, integrated hardware and software silicon IP platform has been announced by CEVA, based on its imaging and vision DSP, the CEVA-XM6. It allows developers to harness the power of neural networks and machine vision for camera-enabled smart devices.
Compared to the earlier CEVA-XM4 intelligent vision DSP, the CEVA-XM6-based vision platform delivers up to eight times higher performance for neural network workloads and up to three times performance improvement across all computer vision kernels. The architecture has been enhanced with a new vector and scalar processing units and enhancements to instruction set, memory bandwidth and direct memory access (DMA).
According to the company, the imaging and vision platform delivers more than 25times the performance/W efficiency and four times faster processing for convolutional neural networks (CNNs) such as AlexNet and GoogLeNet, when compared with leading GPU-based embedded systems for computer vision and deep learning.
Alongside the DSP, the platform includes function-specific accelerators for CNN and image de-warp (for all types of image transformations), the CDNN2 neural network software framework, OpenCV, OpenCL and OpenVX APIs, CEVA-CV computer vision library and a set of optimised, widely used algorithms.
The CEVA-XM6 introduces a vector processor unit (VPU) architecture, ensuring more than 95 per cent MAC utilization; unmatched in the industry today, claims the company. There is also an enhanced parallel scatter-gather memory load mechanism to improve the performance of vision algorithms, including SLAM and depth mapping.
Sliding Window 2.0 is a patented mechanism which takes advantage of pixel overlap in image processing to meet increasing complexity in neural networks. Other features are an optional 32-way SIMD vector floating-point unit that includes the IEEE half precision standard (FP16) and major non-linear operations enhancements. Enhancements to the 3D data processing scheme accelerates CNN performance, and there is a 50 per cent improvement in control code performance, when compared with the CEVA-XM4. Other features area scalar unit which reduces code size, multi-core and system integration support.
The 16bit CDNN accelerator delivers 512 MACs/cycle, to handle today’s complex neural networks. It frees up the 256 MAC units in the DSP, allowing additional computer vision tasks to run in parallel. For wide angle camera applications, such as 360 degree cameras, the image de-warp accelerator supports the ARM Frame Buffer Compression (AFBC) protocol, for best system interoperability.
ISO 26262 active safety compliance supports the needs of next generation ADAS and automated driving solutions for automotive use cases.
The CEVA-XM6 DSP and vision platform components will be available for licensing to lead customers in Q4 of 2016 and for general licensing in Q1 of 2017.