Unmanned Systems Technology 013 | AutonomouStuff Lincoln MKZ | AI systems | Unmanned Underwater Vehicles | Cosworth AG2 UAV twin | AceCore Neo | Maintenance | IDEX 2017 Show report

38 efficient computation engines that can be used as nodes in the network. This has proved especially popular for large learning networks where multiple cards can be combined to provide deep nets. However, there can be a limit to the data pipeline, both to the array of GPUs and between the nodes, because neural nets are multi-dimensional matrices. The GPU programming model has also historically been one of a master-slave from a PC to a graphics processor, and that is not a good development model for neural nets. It creates unnecessary data moves and the need to balance which data is kept in which memories, generating large data movements that need to be coordinated. The x86 and GPU processors are also being linked to field programmable gate array devices (FPGAs) for higher performance. These can implement a neural net in a reprogrammable fabric of logic, and although they take longer to be reconfigured (as a bit stream has to be downloaded into the low-level fabric) they can provide a faster and lower-power implementation of a net. While the large I/O and parallel processing capability of FPGAs is attractive, the skills for development are hard to find and the development time is long. The OpenCL specification addresses this by introducing the concept of pipes that connect data sources and data sinks to the OpenCL kernel without using a master-slave architecture. This combination of processing technologies and development tools is significant. An FPGA with a neural network development framework such as Café means a deep learning system can now be developed much faster than before. FPGAs with large arrays of programmable logic can support up to 20 vision channels operating in parallel, each with its own specific image processing pipeline. These pipelines can be generated from deep learning image processing cores, and connected using April/May 2017 | Unmanned Systems Technology Intel Xeon multicore processors are being used for deep learning applications in driverless cars and UAV designs. The COM Express boards support the latest 16- core Kaby Lake processors with 32 Gbytes of DDR4 memory (Courtesy of Kontron) Nvidia’s PX-2 uses the Pascal GPU processor for the development and prototyping of machine learning algorithms for driverless cars (Courtesy of Nvidia)