CUDA, cuDNN, and DirectX: Why Modern Neural Networks Depend More and More on GPUs

Video surveillance has moved far beyond simply recording footage from a camera. A modern security system is expected not only to store video archives, but also to understand what is happening in the frame: detect people, vehicles, faces, license plates, line crossing, motion in selected zones, a person’s gender and age, alarm events, and many other visual patterns. All of this requires neural networks, and neural networks require serious computing power.

In this release, line-crossing detectors have been optimized for GPU processing, gender and age detectors have been accelerated, additional checks for CUDA, cuDNN, DirectX, and DXCore have been added, and users can now configure how many frames per second are sent to recognition detectors. If the required components are not available on the computer, the system continues working on the CPU. In other words: no GPU, no drama. SmartVision simply falls back to the traditional way and keeps running.

Neural networks perform a huge number of similar mathematical operations. Detecting an object in a frame, analyzing a face, estimating age or gender, or checking whether a person has crossed a virtual line all come down to millions or even billions of calculations: matrix operations, convolutions, comparisons, and feature transformations.

A central processing unit, or CPU, is excellent for general-purpose tasks. It manages the application, operating system, files, network connections, database operations, user interface, and application logic. However, the CPU is not always the best tool for large-scale parallel computing.

A graphics processing unit, or GPU, is built differently. It is designed to process a large number of similar operations at the same time. Originally, this architecture was needed for graphics: calculating millions of pixels, shadows, textures, and lighting effects. Later, it became clear that the same architecture is also ideal for neural networks. Instead of processing pixels for graphics, the GPU can process matrices, image features, and neural network layers.

For video analytics, this became a turning point. When recognition tasks were processed mostly on the CPU, the number of cameras, analysis frequency, and model complexity were severely limited. A system could record video, but full AI-powered analytics across multiple streams would quickly overload the processor. GPUs changed the rules. The most demanding tasks can now be moved to the graphics card, reducing CPU load and increasing processing speed.

CUDA is an NVIDIA technology that allows a GPU to be used not only for displaying images on a screen, but also for general-purpose computing. Simply put, CUDA gives software access to the computing cores of the GPU.

For neural networks, CUDA acts as a foundational layer. It allows software to send computing tasks to the GPU, manage GPU memory, run parallel operations, and receive the results back. Without CUDA, many AI modules cannot use an NVIDIA GPU for acceleration.

In video surveillance, this is especially important. When SmartVision analyzes a camera stream, each frame may pass through one or several detectors. When there are many cameras, the workload increases very quickly. CUDA makes it possible to perform part of this work on the GPU, where parallel processing is much more efficient.

cuDNN is NVIDIA’s specialized library for deep neural networks. If CUDA is the road that gives software access to the GPU’s computing power, cuDNN is the set of professional tools already optimized for traveling that road.

CUDA provides general access to GPU computing. cuDNN provides high-performance ready-made operations commonly used in neural networks, including convolutions, matrix calculations, pooling, normalization, activation functions, and other core building blocks of deep learning.

The difference is simple:

CUDA makes it possible to run computations on the GPU.

cuDNN makes neural network operations run on the GPU as efficiently as possible.

That is why AI modules often require both components. Installing only the graphics card driver is not always enough. The driver allows Windows to recognize the GPU and use it for graphics, but neural network libraries also require additional components such as CUDA Runtime and cuDNN. If the required DLL files are missing, if versions are incompatible, or if the system path is not detected correctly, GPU acceleration may not be enabled, and the system may switch back to CPU processing.

This is why SmartVision 6.3 includes additional checks for CUDA and cuDNN installation. These checks help the system determine more accurately whether the GPU can be used for accelerating neural network modules or whether it is safer to continue running on the CPU.

DirectX is a set of Microsoft technologies for working with graphics, video, multimedia, and hardware acceleration in Windows. Many users think of DirectX mainly as a gaming component, but its role is much broader. DirectX helps applications interact with the Windows graphics subsystem, video drivers, and GPU hardware capabilities.

For video analytics systems, DirectX is important not because a neural network needs to “draw a pretty picture,” but because modern video processing is closely connected with the graphics subsystem. Video display, stream decoding, buffer handling, driver interaction, and graphics device detection may all depend on these system components.

DXCore is a Windows component related to detecting graphics and compute adapters. It helps applications understand which GPUs are available in the system. If DXCore is missing, damaged, or unavailable in a specific Windows version, an application may fail to detect the graphics card correctly or may not gain access to certain hardware acceleration capabilities.

That is why SmartVision 6.3 also checks for DirectX and DXCore. This is especially important across different versions of Windows, where system components may vary. On one computer, all required modules may already be installed. On another, a driver update or additional component installation may be required. On a third system, everything may work only on the CPU because the GPU is outdated, the driver is incompatible, or the required libraries are missing.

In the past, video analytics was much more limited. Video surveillance systems mostly recorded archives, displayed live camera feeds, and reacted to basic motion. Those algorithms were relatively lightweight: compare pixel changes, detect motion, and record an event.

True intelligent analytics requires a much higher level of computing power. Detecting a person in a frame, distinguishing a vehicle from an animal, recognizing a face, estimating age and gender, reading a license plate, or tracking line crossing is far more complex than simple motion detection.

Performance used to be the main limitation. A CPU could handle one or several tasks, but as the number of cameras increased, the load became too high. Hardware cost was another problem: powerful servers were expensive, while ordinary computers quickly reached their limits. Software compatibility was also a major obstacle. Stable GPU acceleration requires the right drivers, CUDA, cuDNN, compatible library versions, operating system support, and proper implementation inside the application itself.

The development of GPUs, neural network libraries, and hardware acceleration made modern video analytics practical and widely accessible. Tasks that once required expensive servers can now often run on a regular workstation equipped with a suitable NVIDIA graphics card. It is not magic, although from the outside it may look close. It is simply math finally being sent to the place where it belongs.

SmartVision 6.3 uses GPU capabilities to accelerate neural network detectors and make the system more flexible. In this version, line-crossing detectors have been optimized for GPU processing. This is important for security scenarios where the system must quickly detect movement across a defined boundary: entering a restricted area, leaving a room, crossing a perimeter, or moving through a control line.

Gender and age detectors have also been optimized for GPU processing. These functions require facial and human feature analysis, so they can create noticeable load when processing many frames. Using the GPU helps distribute these calculations more efficiently and reduce CPU usage.

SmartVision also adds a setting that controls how many frames per second are sent to recognition detectors. This is an important practical feature. It is not always necessary to analyze every frame of a video stream. Some tasks require only a few analyzed frames per second, while others need a higher recognition frequency. This new setting allows users to balance response speed, analytics accuracy, and computer workload.

If the computer has a compatible NVIDIA graphics card and CUDA, cuDNN, DirectX, and DXCore are installed correctly, SmartVision can use hardware acceleration. If the required components are not available, the system continues working on the CPU. This approach makes the software compatible with a wide range of configurations, from standard office PCs to high-performance GPU workstations.

In addition, SmartVision 6.3 includes improvements to the quick Alerts window, making it more convenient for monitoring events, notifications, and alarm triggers. The interface has also been refined, and translations into multiple languages have been updated.

Most video surveillance users do not want to manually investigate where CUDA is installed, which version of cuDNN is available, or why one DLL file is visible to the application while another one is not. They need the cameras to work, events to be detected quickly, archives to be recorded reliably, and the interface to stay clean instead of turning into a museum of error messages.

SmartVision 6.3 moves in exactly that direction. The system checks the required components for GPU acceleration, uses the graphics card where possible, and keeps working on the CPU when hardware acceleration is unavailable. This is the right logic for real-world video surveillance: maximum performance on modern computers and reliable compatibility with standard systems.

CUDA and cuDNN give neural networks access to the power of NVIDIA GPUs. DirectX and DXCore help Windows and the application work correctly with the graphics subsystem and video devices. Together, these components form the foundation of modern video analytics, where a camera no longer simply “watches,” but helps the system understand what is happening in the frame.

SmartVision 6.3 applies this approach in practice: it accelerates detectors, adds component checks, gives users control over frame analysis frequency, and maintains stable operation even without a GPU. This is how a modern video surveillance system should evolve: faster, smarter, and without unnecessary ritual dances around system libraries, though in IT, the tambourine is never fully retired.