Performance IP processor core neural networks VeriSilicon Vivante VIP8000 more than 3 trillion operations per second MAC

The company VeriSilicon introduces IP — Vivante VIP8000. According to VeriSilicon, Vivante VIP8000 — programmable and scalable processor for machine vision systems and artificial intelligence. Its performance exceeds 3 trillion operations MAC (Multiply-Accumulate Computing) per second. While energy efficiency is 1.5 billion MAC / second per 1 mW of power consumption. Additionally, being designed for the production 16FF process technology, the processor VIP8000 is the smallest product of its kind in the industry.

Configuration Vivante VIP8000 includes multi-threaded processing unit, the block neural networks and block a generic cache memory. The processor can directly load the neural network, generated using popular tools, including Caffe and TensorFlow. Using the library OpenVX, these networks can be integrated with other components of machine vision systems. The processor supports all popular models of networks, including AlexNet, GoogleNet, ResNet, VGG, Faster-RCNN, Yolo, SSD, FCN and SegNet, and layers. Programmability provides support for patterns and layers that may appear in the future. For programming the CPU can use OpenCL or OpenVX.

Source: VeriSilicon


(Visited 44 times, 1 visits today)
No tags for this post.