FPGAs – Ideal for Deep Learning
Due to their architecture to process data with high parallelism, FPGAs are predestined for image processing as well as for the implementation and execution of neural networks. They excel with guaranteed robust image acquisition and — in comparison to CPUs and GPUs — high processing power, frame rate and bandwidth that allow CNNs on FPGAs to classify with high data throughput, fulfilling the demands of inline inspection in particular. The FPGA enables processing of image data directly on a frame grabber or an embedded vision device — from image acquisition to analysis result — without burdening the CPU, which is important for CPU-intensive applications such as CNNs. Thus, smaller PCs without GPUs can be used, reducing overall system costs.
Programmable CNN ready frame grabber
FPGAs’ energy efficiency in the industrial temperature range is 10x higher than that of GPUs. This is ideal for embedded devices, clearly expanding the field of use with regard to Industry 4.0 as well as drones and autonomous driving. Hard- and software from Silicon Software supports SoCs and single board computers (SBCs) alongside FPGAs. For an FPGA, the shift into the fixed point area means that the resources can be used for larger network architectures or for higher data throughput. Moreover, effective image preprocessing that reduces data enables use of smaller networks or FPGAs. These often suffice for simple classification tasks with few characteristics.
Users are in a position to program deep learning applications on their own on the FPGA — in little time using drag & drop with no hardware programming skills, thanks to the VisualApplets graphical programming environment. This is equally valid for the integration of image processing peripherals such as actuators and sensors via real-time signal processing. Users can simply continue to use their existing image processing system. The long-term availability of FPGAs, frame grabbers, and VisualApplets guarantees a high level of investment security.