Omnitek DPU for Embedded / Edge Applications
The highly efficient design of the Omnitek DPU enables it to be configured either to deliver very high performance or, more typically for embedded applications, the lowest device cost and power consumption for a given performance requirement. For example, the DSP array size can be configured to optimise the performance, cost and power for a particular application. Embedded applications will often use an SoC FPGA with Arm subsystem but may use an FPGA; the Omnitek DPU can be used in any of these platforms.
Embedded applications include surveillance (both static cameras and drones), driver assist systems, facial recognition systems, etc. Omnitek is able to create a complete system design around the DPU for such applications using its extensive IP library which has been optimised for FPGAs. The diagram shows an example augmented reality headset design which combines the DPU with IP for MIPI, Camera ISE, Image Warping and embedded Display Ports. The full range of IP available can be found here.
Once configured, the DPU can be software programmed by the user to refine the detailed architecture of the selected DNN. For example, in the case of a CNN, the programmer can select an existing design such as GoogLeNet or YOLO, or define a fully customised CNN architecture.
The re-configurability of the Omnitek DPU for embedded applications also provides a significant level of future-proofing as new and improved network topologies become available. To achieve this remotely, the system simply needs to accessible over a network. This is a significant advantage compared with dedicated hardware platforms such as ASICs or ASSPs, which are effectively hard-wired to a single topology that quickly becomes obsolete as a result of rapid advances in DNN research.