openvino INT8 FP16 FP32 数据类型性能对比

先说结论:

  • 在CPU上,INT8略快;在GPU上,FP16最快;
  • 使用GPU比使用CPU节约80%左右的CPU开销;
  • 使用GPU会比使用CPU占用更多的内存,不知道是不是因为显卡没有现存,共享内存导致的;
  • 模型加载时间差异很大,使用GPU可以大大节省CPU资源,但初始化加载模型的过程较慢;

测试环境:intel core i5 8500
应用:human_pose_estimation_demo
模型:human-pose-estimation-0001
数据:MOT16/test/MOT16-06

耗时(ms)CPUGPU
FP16-INT8124559179275
FP1612878881001.2
FP32127422103820

模型加载时间:

  • CPU FP16-INT8:小于2s;
  • GPU FP16:大约53s;

CPU占用率差异:

  • CPU FP16-INT8:290%
  • GPU FP16:60%

内存占用差异:

  • CPU FP16-INT8:VIRT:1.18G,RES:136M
  • GPU FP16:VIRT:1.59G,RES:528M

日志:
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -m intel/human-pose-estimation-0001/FP16-INT8/human-pose-estimation-0001.xml
InferenceEngine: 0x7f3ae5583030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 124559

Execution successful
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -m intel/human-pose-estimation-0001/FP16/human-pose-estimation-0001.xml
InferenceEngine: 0x7fbb0eaa2030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 128788

Execution successful
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -m intel/human-pose-estimation-0001/FP32/human-pose-estimation-0001.xml
InferenceEngine: 0x7f9136f9f030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 127422

Execution successful
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -d GPU -m intel/human-pose-estimation-0001/FP32/human-pose-estimation-0001.xml
InferenceEngine: 0x7f1a2209e030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 103820

Execution successful
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -d GPU -m intel/human-pose-estimation-0001/FP16/human-pose-estimation-0001.xml
InferenceEngine: 0x7fe9f13a0030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 81001.2

Execution successful
root@openpose:/openpose_lite# /root/omz_demos_build/intel64/Release/human_pose_estimation_demo -i MOT16/test/MOT16-06/img1/%6d.jpg -no_show -d GPU -m intel/human-pose-estimation-0001/FP16-INT8/human-pose-estimation-0001.xml
InferenceEngine: 0x7fe0ce749030
Parsing input parameters
Reshape needed
To close the application, press ‘CTRL+C’ here
Total Inference time: 179275

Execution successful

  • 1
    点赞
  • 4
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

飞花丝雨

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值