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

an AXI streaming interface. This is a standard interface and is used on all the IP modules to reduce development time, allowing for an easier upgrade for future designs. The latest 17 nm silicon chip processing technology is allowing these large FPGA devices, as well as the latest GPU array processors, to have sufficiently low power consumption to be used in UAVs and driverless cars. UAV designs are of course determined by the power available from the battery, while automotive designs are more about the power dissipation and the temperature of the chip and the board. Early versions of the GPU cards had to be water-cooled because they generated so much heat, but devices using 10 nm technology can be air-cooled and provide even higher performance for the AI algorithms. While image classification convolutional neural nets lend themselves to SIMD (single instruction, multiple data) architectures such as those in GPUs, other elements of AI such as multi-agent processing need multiple instruction, multiple data (MIMD) structures to build simulations with 100 different instances of the environment going forward and then choose the safest. This is driving the development of a new class of hardware optimised for neural nets. One approach here is to look at a neural network as a ‘graph’. A neuron forms the vertex of the graph, and the edges store data as a weighting that determines the likelihood of connecting to another neuron. A computational graph then looks like a neural network based on a compute function, similar to a probability graph. All the machine learning frameworks such as TensorFlow and MixNet are graph languages. Using this graph approach allows a machine learning system to AI systems | Focus Devices that use 10 nm technology can be air-cooled and provide even higher performance for AI algorithms