The company continues to see a lot of demand for smart sensing systems that combine deep learning with other computer vision and video processing techniques such as SLAM or structure from motion, wide-angle lens correction, and video compression. It claims to be the only company that can run all these tasks on a single unified processing architecture, simplifying SOC design and integration while easing software design and reducing unused dark silicon.
“We’ve quietly been working on our deep learning solution together with a few select customers for quite some time and are now ready to announce this exciting new technology to the broader market,” says Hans-Joachim Stolberg, CEO at videantis.
“To efficiently run deep convolutional nets in real-time requires new performance levels and careful optimization, which we’ve addressed with both a new processor architecture and a new optimization tool. Compared to other solutions on the market, we took great care to create an architecture that truly processes all layers of CNNs on a single architecture rather than adding standalone accelerators where the performance breaks on the data transfers in between.”
"Using some clever design features, the v-MP6000UDX architecture we’re announcing today increases throughput on key neural network implementations by roughly 3 orders of magnitude, while remaining extremely low power and compatible with our v-MP4000HDX architecture. This compatibility ensures a seamless upgrade path for our customers toward adding deep learning capabilities to their systems, without having to rewrite the computer vision software they’ve already developed for our architecture”, Stolberg continued.
The v-MP6000UDX processor architecture includes an extended instruction set optimized for running convolutional neural nets, increases the multiply-accumulate throughput per core eightfold to 64 MACs per core, and extends the number of cores from typically 8 to up to 256.
Alongside the new architecture, videantis also announced v-CNNDesigner, a new tool that enables easy porting of neural networks that have been designed and trained using frameworks such as TensorFlow or Caffe.