As discussed in the System Design chapter, the vision system processing consists of two main components:

• Model Development phase—where Deep Learning tools analyze the labeled, training image set, and the network is trained. This phase is presented using the Cognex VisionPro ViDi Deep Learning Windows application.
• Runtime phase—where the tools operate on input images and produce markings based on their training data. This phase is also called the Deployment phase and is generated by the Cognex Designer application and runs as a Windows executable on a far edge platform.

  Each section below discusses the sizing guidelines for the two distinct phases, providing guidance on platform selection, CPU/memory benchmarks, as well as GPU considerations. Where application consolidation is an option (training + runtime), the aggregate sizing recommendations are provided.

Performance for the ViDi Deep Learning application, specifically the training and model development, have several dependencies outside of the characteristics of the hardware platform (CPU, Memory, GPU). With total training time for model development being the core measure, the design pattern, the tools chosen, and the complexity of the images being analyzed all factor into how long the model takes for training. The following are some tuning parameters that can assist in performance optimization during the training phase:

• Utilize a small number of tools—Using a single tool instance to examine multiple image features (red Analyze tool to find multiple defects or green Classify tool to make multiple classifications).
• Avoid image conversions—Image sampling for tool operation uses raster (bitmap) processing. Attempt to use single-channel grey-scale images.
• Reducing the amount of processed data improves processing speed. This may be accomplished by:
  • Using a smaller ROI
  • Using a mask
  • Using fewer image channels
• Multi-threading—On systems with multiple GPUs, processing multiple streams concurrently allows tools to execute in parallel, increasing output.
• Tool parameters directly affect tool execution speed. The following tool parameters can be modified to increase performance. There is a balance between tool speed, accuracy, and robustness:
  • Feature size—At run time, ViDi tools need to sample the entire input image. The feature size determines the number of samples required for a given image size. The larger the feature size, the fewer samples needed. Larger feature sizes may cause the tool to miss features or defects.
  • Sampling density—Similarly to feature size, the sampling density determines the number of samples required for a given image size. Increasing the density increases the processing time linearly.

The development environment for training of the ML model must take a number of factors into consideration. Memory and storage considerations come into play when looking at the overall model development environment. An individual workspace is used for model development, and each workspace within Cognex Designer uses its own memory footprint. If the development team is working on multiple models for products simultaneously, then the platform resources need to accommodate the processing and memory requirements of doing so.

Defining a small, medium, and large configuration for a vision application is not as straight forward as for other types of applications, since some applications require extremely high inspection speed while using a high number of cameras as part of the same inspection process. Therefore, vision application integrators play a vital role in selecting the appropriate hardware configuration for the particular application at the time of development. However, if extreme cases are left out of the equation, most vision applications will successfully run on the hardware specifications in the following table.

The Cognex Designer runtime, when deployed as a stand-alone executable (with no development environment), can be viewed as a more versatile component in the overall Vision System architecture. Located at the far edge on the industrial factory floor, the ruggedized nature of these hardened compute platforms provides the connectivity requirements to GigE cameras for capturing product images.

• The Dell XR4000 has the additional capacity for a NVIDIA GPU. This provides enhanced performance for executing ML models developed with Deep Learning tools.
• Due to being GPU enabled, the XR4000 can run ViDi Deep Learning applications as well asVisionPro applications where GPU is not required . The system can be used to run small and medium-sized vision applications.
• As a MIL/NEBS tested server, the XR4000 can operate in temperatures from –5°C to 55°C, withstanding most levels of shock, vibration, and dust deployments.

  Table 24. Cognex runtime sizing guidelines

  Platform	Small/Medium	Large
  Dell XR4000 NVIDIA GPU	8 CPU 32 GB RAM 3 USB ports 3 Gig-E interfaces	16 CPU 64 GB RAM 3 USB ports 3 Gig-E interfaces
  Dell Edge Gateway 5200	8 CPU 32 GB RAM 4 USB ports 3 Gig-E interfaces	Not Recommended

• The Dell Edge Gateway 5200 provides up to eight CPUs and SSD capacity in a compact ruggedized design. Although not GPU capable, for use-cases that do not require high performance the EGW-5200 can accept multiple Gig-E camera inputs for image capture and runtime processing.
• As a far edge designed platform, the EGW-5200 has the utility to accommodate multiple functions.

In our testing we observed the following scenarios:

• Cognex Designer runtime without ViDi Deep Learning tools—In this scenario, the presence of a GPU does not impact performance. The application performance relies on the CPU frequency and number of cores. In a scenario where a single image acquisition task is used, the performance relies on the frequency of the CPU. The higher the frequency, the higher the processed output within a given time interval (images/pixels inspected per second). If two or more image acquisition tasks are being executed asynchronously, the number of CPU cores running at the same frequency as before must increase accordingly to obtain the same output performance. If using multiple inspection tasks and running them concurrently on the same CPU core, the output of each task is divided accordingly.
• Cognex Designer runtime with ViDi DeepLearning tools—In this scenario, the GPU plays a more prevalent role, and the application output (images/pixels inspected per second) will greatly depend on the GPU’s frequency and number of CUDA Cores available. Increasing the CUDA Cores count of the GPU being used or using multiple GPU’s will improve the application’s output accordingly. If the vision application is using a combination of traditional vision tools and Deep Learning tools, the CPU will also play a major role in the application’s output, depending on the combinations of tools being utilized and the overall design of the application.

  For more information, see Cognex VisionPro ViDi Help - Deployment PC Requirements - Documentation | Cognex

• Some VisionPro vision tools are multi-thread aware, and can be programmatically enabled to operate on threads. This can be accessed using a script, such as the following example (the script has a Using statement of Cognex.VisionPro):

  CogVisionToolMultiThreading.Enable = true;
  CogVisionToolMultiThreading.ThreadCountMode = CogVisionToolMultiThreadingThreadCountModeConstants.HardwareDefined;
  

  This option enables multi-threading on a VisionPro vision tool level (for example, a CogBlobTool). In this way, VisionPro applications can be designed for increased performance by allowing parts of the inspection to be executed in parallel.

• Cognex Designer supports parallel processing with multi-core CPUs through the use of the parallel region, which allows the processing of intensive applications to be optimized for fast execution. The parallel region is constrained by the number of processor cores on the PC, both for development and deployment. For instance, if your development/deployment PC only has two processor cores, you should only have two VisionPro Tool Blocks within a parallel region. If the parallel region contained three VisionPro Tool Blocks, the three blocks would not be executed simultaneously, because the PC only has two processor cores.
• In order to take advantage of CPU multi-threading in vision applications using Cognex VisionPro APIs and Cognex Designer, the application developer must implement the vision tools with multi-threaded operation in mind.

• On systems with multiple GPUs, processing multiple streams concurrently allows tools to execute in parallel, increasing throughput. Use the --max-process-count command line argument to enable multiple threads to access a single GPU.
• On single-GPU systems, you can configure the system to allow multiple processes to make use of the same GPU. This allows a higher GPU occupancy and can improve throughput, although tool execution time will increase. To enable multi-process GPU access for a runtime application using the local control's GlobalConfig() method: control.GlobalConfig( "max_process_count=2").

Since vision applications are highly customizable, the performance of the overall application is dependent on several aspects:

• The number of acquisition devices (cameras) used in the application and the resolution of the captured images
• The percentage of the pixels out of the entire picture being analyzed by the vision tools (feature size)
• The number of vision tools used, as well as the settings used for those tools (sampling density)
• The knowledge of the application design team, in their designing and in their use of best practices in computer vision applications

The hardware required for a specific application must be chosen at the time when a working model of the vision application exists, and initial performance can be tested. Running a proof-of-concept (POC) or leveraging benchmarks from similar application deployments can also assist in sizing the platform requirements.