The benchmark shows more lower speed in int8 than float. It cann't satisfy our expection

[aqsc]

expectation

https://github.com/Tencent/ncnn/tree/master/benchmark From the benchmark website, it shows the inference in int8 quantization cannot be faster than float32 in arm platform and other platforms. It doesn't satisfy our expection.

[wzyforgit]

Because float32 is so well optimized than int8

[nihui]

The data in the benchmark readme is likely out of date The new version optimizes the inference performance of int8 Welcome to update and test again. If there are any problems, please reopen the issue.