Hardware requirements for neural analytics operation

When you use neural analytics (see General information on neural analytics), the hardware requirements are the following:

1. Take into account the limitations of NVIDIA SDK. The neural analytics can operate on the GPU only in the operating systems specified in the Operating system requirements page.
2. VPU requirements (Intel NCS):
   1. Connect a device to the computer USB port.
   2. Make sure that the Windows system identifies a device as Movidius, Myriad X, or VSC Loopback Device.
   3. Use it on the computers that meet the Hardware requirements.
   4. Apply no more than one Intel NCS device on the server.
3. Supported CPUs and requirements:
   

   CPU type	Supported generations/models	Key requirements and conditions
   Intel® Core™	6th generation (Skylake) and higher	Support for the AVX2 or AVX512 instruction set (see ark.intel.com) We don't guarantee the detector operation on CPUs lower than 6th generation
   Intel® Xeon®	Families v5 (Broadwell) and v6 (Skylake)	Support for the AVX2 or AVX512 instruction set (see ark.intel.com). CPU compatibility with the OpenVINO toolkit version that you use (see software.intel.com)
   Intel VPU devices	Movidius™ Neural Compute Stick, Neural Compute Stick 2, Vision Accelerator Design with Intel® Movidius™ VPUs	Support from the OpenVINO toolkit version that you use (see software.intel.com)

4. Requirements for NVIDIA GPU: 
   

   Parameter	Minimum/recommended requirements
   GPU model	NVIDIA GeForce GTX 1050 and later (for Detector Pack 3.14–3.15.1) NVIDIA GeForce GTX 1650 and later (for Detector Pack 3.15.2 and later)
   Video memory space	At least 2 GB. For the accurate calculation, we recommend using AxxonSoft Platform Calculator (see AxxonSoft Platform Calculator documentation)
   CUDA version	11.1–11.8 (for Detector Pack 3.14–3.15.1) 12.8.1 (for Detector Pack 3.15.2 and later)
   Compute Capability	6.0–9.0 (for Detector Pack 3.14–3.15.1) 7.5–12.0 (for Detector Pack 3.15.2 and later) You can check the version here or on the manufacturer's website
   Supported architectures	Kepler (late), Maxwell, Pascal, Volta, Turing, Ampere, Ada Lovelace (partially) (for Detector Pack 3.14–3.15.1) Turing, Ampere, Ada Lovelace (partially), Hopper, and Blackwell (for Detector Pack 3.15.2 and later) In Windows OS, starting with the Detector Pack 3.15.2 version, the support for Hopper and Blackwell is added for the following detectors: Neural tracker Neural counter Human pose detector Equipment detector Fire detector Smoke detector Neural classifier Barcode detector Water level detector  Stopped object detector Privacy masking Person-based privacy masking
   Drivers	We recommend installing the latest driver version from the official website
   Configuration	You can use up to four GPUs on one server

5. For the correct operation, each detector must meet individual specific requirements (see Detectors) to:
   • Hardware platform,
   • Video stream and scene,
   • Image characteristics.