Xilinx / video-sdk

https://xilinx.github.io/video-sdk
Other
30 stars 14 forks source link

Decode 1080p@5fps #72

Closed kevinzezel closed 1 year ago

kevinzezel commented 1 year ago

Hi,

I saw in the documentation that each Alveo U30 card can decode 16 streams at 1080p@30fps. If I lower the frame rate of the RTSP cameras to 5 FPS, can I increase the amount of stream per camera?

Is it correct to say: 16x 1080p@30fps 96x 1080p@5fps

How can I monitor card usage? Similar to htop for CPU and nvtop for GPU.

Regards, Kevin

NastoohX commented 1 year ago

Hi, Thank you for your question. Although, the card allows for linear increase in density, as an inverse function of FPS; however, at higher densities, other factors including PCIe traffic and control software overhead may show a diminishing return. You need to test for a stable density, in your specific use case. We report on typical use-cases here: https://xilinx.github.io/video-sdk/v3.0/specs_and_features.html#performance-tables. To get performance numbers out of U30, use xrmadm /opt/xilinx/xrm/test/list_cmd.json command. For example if you were to decode a 4kp60 content, the decoder should report full utilisation: "cu_3": { "cuId ": "3", "cuType ": "IP Kernel", "kernelName ": "decoder", "kernelAlias ": "DECODER_MPSOC", "instanceName ": "decoder_1", "cuName ": "decoder:decoder_1", "kernelPlugin ": "/opt/xilinx/xma_plugins/libvcu-xma-dec-plg.so", "maxCapacity ": "497664000", "numChanInuse ": "1", "usedLoad ": "1000000 of 1000000", }, , whereas, if you were to decode a 4kp5 content, it will report 1/12th of utilisation: "cu_3": { "cuId ": "3", "cuType ": "IP Kernel", "kernelName ": "decoder", "kernelAlias ": "DECODER_MPSOC", "instanceName ": "decoder_1", "cuName ": "decoder:decoder_1", "kernelPlugin ": "/opt/xilinx/xma_plugins/libvcu-xma-dec-plg.so", "maxCapacity ": "497664000", "numChanInuse ": "1", "usedLoad ": "83333 of 1000000", }, Hope this helps. Cheers,

kevinzezel commented 1 year ago

Thank you!!

NastoohX commented 1 year ago

Great. I'll close this ticket for now, but feel free to reopen it or issue a new one. Cheers,