search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

Hardware setup for multistream on v100 for 300 cameras #6957

Open AGpix opened 3 years ago

AGpix commented 3 years ago

Hello I need to process around 300 cameras streams in real time. Size of Yolov4-tiny-3l network I am using is 544X544. I need to estimate how many streams I can process on one v100. I calculated V100 could probably process 70/80 streams is my calculation correct. Also what is the cpu and RAM requirement for this. I haven't used tkdnn yet so would to know for without tkdnn.

I can't afford to go wrong on this as its an expensive rig. Thanks

stephanecharette commented 3 years ago

I don't have an answer for you. I did recently figure out how much memory is required to train a network: https://www.ccoderun.ca/programming/2020-09-25_Darknet_FAQ/#memory_consuption

What you need to figure out is the memory requirements when batch=1 and subdivisions=1, cause those tables are for when batch=64. You also don't state what "real time" means to you. Are you talking about having a full 30 FPS? Are you planning on having a darknet processes per stream? Or will you multiplex several streams per darknet process?

AGpix commented 3 years ago

@stephanecharette, sorry forgot to mention for me real time is 12 to 15 fps . I can do threading per darknet process. But according to @AlexeyAB its better to run one darknet process for each stream. So latter is preferable.

AGpix commented 3 years ago

@AlexeyAB please could you suggest. I have looked at earlier posts you have made some recommendations around GTX2080 but not sure how it translates to v100