AI accelerator cards for AIoT edge computing

The topic of artificial intelligence is on everyone's lips. In industry, AI-supported applications are playing an increasing role.
In production, optical recognition is used in manufacturing processes to detect production defects, for example. Industrial PC systems, or so-called inference systems, are used to make decisions about the condition of production products based on images.
These inference systems use pre-trained data sets, so-called training models, which have been created in high-performance computers.

While these training servers usually use multiple graphics cards to classify the model as quickly as possible, the use of graphics cards at Edge is not practical or possible for a variety of reasons. Among other things, requirements such as the physical size of the inference system, power consumption and the price-performance ratio play a role. These requirements cannot be met with GPU-based systems. The Mustang AI accelerator card series offers new possibilities for edge computing or Artificial Intelligence of Things (AIoT). Two different Mustang series are available to realize flexible and scalable AI applications at the edge.

VPU accelerator cards

The VPU based Mustang series uses Intel^® Movidius^TM Myriad^TM X MA2485 visual processing units. These VPUs are specifically designed for AI applications in the vision domain through the combination of their neural network with 16 SHAVE cores and their neural compute engine.
In addition to a variety of implemented hardware features for image processing, each VPU includes a new Stereo Depth Block feature capable of processing two streams at 720P resolution at 180 Hz. Native FP16 calculations or 8bit fixed point calculations but also the tuneable ISP pipeline and the hardware-based encoding also allow sophisticated image or video processing with a resolution of up to 4K.
Each individual VPU can be assigned a different DL topology. The reason for this is the multi-channel capability of the VPUs, which enables simultaneous execution of calculations.
This allows different applications such as object recognition or image and video classification to be executed simultaneously.