MyVMS uses GPU acceleration for AI video analytics. GPU requirements depend on the operating system, enabled analytics and the number of analytics channels.
GPU compatibility and GPU performance are different requirements. A compatible GPU may run MyVMS AI analytics but still provide insufficient performance for a large analytics workload.
For new systems using AI analytics, we currently recommend NVIDIA GeForce RTX 3050 or higher.
MyVMS supports NVIDIA, AMD Radeon and Intel GPUs on Windows.
AI acceleration on Windows uses DirectML. NVIDIA CUDA support is not required.
MyVMS currently supports NVIDIA GPUs only for AI acceleration on Linux.
For new installations, NVIDIA GeForce RTX 3050 or higher is currently recommended.
Different AI features require different amounts of dedicated GPU memory (VRAM). The values below can be used to estimate the GPU memory required for your planned analytics workload.
For example, 10 Object Detection channels and 3 Face Analytics channels require approximately 2.4 GB of GPU memory.
Use these estimates to calculate the approximate VRAM required for your analytics configuration and compare it with the available memory of the selected GPU.
Actual GPU memory consumption may vary depending on the GPU model, operating system, MyVMS version and AI model version. These values are estimates, not fixed hardware requirements.
GPU memory determines whether the required AI workload can fit into the available graphics memory, but it does not by itself determine processing performance.
GPU architecture, computational performance, memory bandwidth and the complete system configuration also affect the number of analytics channels that can be processed in real time.
Because hardware specifications provide only theoretical performance, the maximum practical number of analytics channels can be determined accurately only through real-world testing.
First estimate the GPU memory required for the planned analytics channels using the values above. Then select a compatible GPU with sufficient VRAM and performance for the expected workload.
Meeting the minimum compatibility requirements does not guarantee sufficient performance for every AI configuration.