Test platform
Intel i5-7440 hq,
Win10 1709
GPU driver 26.20.100.7812 release date 1/21/2020
Use OV bring their own benchmark tests, test case 2
Minimum delay (as soon as possible processing a single picture)
Benchmark_app. Exe - nireq 1 - nstreams 1-1 - b m mobilenetV3 \ [FP16 | FP32] \ v3 - large_224_1. 0 _float. XML - d [GPU | CPU]
The maximum throughput throughput (as much as possible with pictures)
Benchmark_app. Exe - m mobilenetV3 \ [FP16 | FP32] \ v3 - large_224_1. 0 _float. XML - d [GPU | CPU]
Test model mobilenetv3 - large_224_1. 0 _float
The test model sqeezenet1.1
The feeling of personal use OpenVINO 2020
Reasoning part of the CPU speed not too obvious, but the GPU reasoning part should be change a lot, tested mobilenetv3 and squeezenet1.1, feeling of part of the underlying GPU reasoning clDNN really spent some effort to optimize,
Using OpenVINO 2020 version of the GPU to do reasoning, if found error when loading network or reasoning, can consider to upgrade the GPU driver, my original version is in March 2019, do FP16 reasoning when loading error, it is good to update the latest driver:)
Using integrated graphics do reasoning students can try the 2020 version, it may be a surprise:)
OpenVINO2020r2 version to support the GPU INT8 reasoning, but whether in my old generation of CPU HD630 running on July is unclear, INT8 conversion tools into a POT, is still in groping...