--- tags: handoverOK, darknet_test --- # AI_darknet Measuring performance ### 5204 Enviorment > 目前相關資源都放在 5204上~這台配有1T 的資料碟與256G ssd系統碟 > 你可存取5204 ~透過ssh scp SAMBA > SAMBA 和 SSH 帳密皆為aewin/aewin ! > 目前5204 IP為 192.168.99.90 ``` ssh aewin@192.168.99.90 or scp aewin@192.168.99.90:/home/aewin/test_file ./your_pc ``` # darknet source code ``` /home/aewin/work/darknet ``` # Darknet/yolo3 neural network (customer provides) ``` yolov3-tiny_100000.weights yolov3-tiny.cfg ``` # test picture 1280*720 ``` 213.jpg 1280*720 ``` ![](https://i.imgur.com/DgH4HnS.png) - How many GFlops(gigaFLOPS:Floating-point) is required for using yolov3-tiny_100000.weights and yolov3-tiny.cfg to inference single picture (720P)? ``` 5.494BFLOPS ``` # GPU Performance test for NVIDIA Tesla P100 PCIe 16 GB on YOLO3 ![](https://i.imgur.com/7Y2qpbY.png) https://www.techpowerup.com/gpu-specs/tesla-p100-pcie-16-gb.c2888 ![](https://i.imgur.com/1EFMoEB.png) ``` 5.494BFLOPS / 0.005357 sec = 1025.6 Gflpos ``` # CPU Performance test for Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz on YOLO3 ![](https://i.imgur.com/pTZlTs6.png) ``` 5.494BFLOPS / 2.71 sec = 2.02 Gflpos ``` # CPU Performance test for Intel(R) Atom(TM) Processor E3930 @ 1.30GHz on YOLO3 ![](https://i.imgur.com/pTZlTs6.png) ``` 5.494BFLOPS / 4.049135 sec = 1.356 Gflpos ``` ----------------------------------- # Camera environment #### 3D 30fps : ![](https://i.imgur.com/AcV7KXH.png) #### 2D ``` 576* 576* 5fps ``` #### calculation 1: ``` 900camera * 5fps = 4500 150camera * 30fps * 2 (3D) = 9000 9000 + 4500 = 13500 frames / sec 算力： 13500 * 5.494BFLOPS = 74169 Gflops = 74 Tflops ``` # How many GPU card to be needed if the system could process inferencing for 13500 frames per second based on YoLOV3 ==NVIDIA GeForce GTX 1080 Ti== ![](https://i.imgur.com/08KcxjE.png) ``` 74 Tflops / 11.3 = 6.5486 ``` ==NVIDIA GeForce GTX 980== ![](https://i.imgur.com/mbqnzM9.png) ``` 74 Tflops / 4.9 = 15 ``` ==NVIDIA Tesla P100== ![](https://i.imgur.com/GNv73PM.png) ``` 74 Tflops / 9.5 = 7.7 ``` # How many 3D and 2D camera we can set up for FPGA of 1 TFLOPS computing power to do inference #### 3D/2D computing power per frame : ``` 1920*1080=2073600 , 12.35 Gflops (3D) 1280*720=921600 , 5.494 Gflops (3D) 576*576=331776 , 2.03 Gflops (2D) 648*480=307200 , 1.83 Gflops (3D depth) ``` ### 3D video 5.494 Gflops*30 fps = 164.82 Gflops ### 3D Depth 1.83Gflops * 30 fps = 54.9 Gflops ### 3D total (Video + Depth) 164.82 + 54.9 = 220 Gflops ### 2D video ,5fps 2.03 Gflops * 5 fps = 10 Gflops ### 2D Video, 25fps 2.03 Gflops * 25 fps = 50 Gflops ## so, How many 2D cameras could be set up when we have set 3D camera of 1 unit for 1 TFLOPS computing power FPGA ? ``` 1000 Gflops - (1* 220 Gflops) = 780 Gflops 780 Gflops / 10 Gflops = 78 cameras for 2D 5fps 780 Gflops / 50 Gflops = 15 cameras for 2D 25fps ``` ### 3D camera of 2 unit .... ``` 1000 Gflops - (2* 220 Gflops) = 560 560 Gflops / 10 Gflops = 56 cameras for 2D 5fps 560 Gflops / 50 Gflops = 11 cameras for 2D 25fps ``` ### 3D camera of 3 unit ... ``` 1000 Gflops - (3* 220 Gflops) = 340 340 Gflops/ 10 Gflops = 34 cameras for 2D 5fps 340 Gflops/ 50 Gflops = 6 cameras for 2D 25fps ``` ### 3D camera of 4 unit ... ``` 1000 Gflops - (4* 220 Gflops) = 120 120 Gflops/ 10 Gflops = 12 cameras for 2D 5fps 120 Gflops/ 50 Gflops = 2 cameras for 2D 25fps ``` PS: https://groups.google.com/forum/#!topic/darknet/W2jKc3ggUfg