[Performance]: Why GPU run slower than CPU on i7 7700 ?

[ke__xu]

Hi All,

Why? Why? Why?

When use OpenVINO to run ResNet_50_cloth_iter_30000 model：

1 My test env:

Win 10，Intel i7-7700 3.6 GHz，8G Memory；OpenVino ver 1.4

Intel i7-7700 ：https://www.intel.cn/content/www/cn/zh/products/processors/core/i7-processors/i7-7700.html

Intel i7-7700 GPU：英特尔® 核芯显卡 630

2 Test Tesult：

CPU ：22.5431 ms； CPU ：32.396ms

[more info] My Code:

        // --------------------------- 1. Load Plugin for inference engine -------------------------------------
        //std::string pluginName {"GPU"};
        std::string pluginName {"CPU"};
        if (MY_DEBUG) std::cout << "[ plugin ] Loading plugin=" << pluginName << std::endl;
         InferencePlugin plugin = PluginDispatcher({"../../../lib/intel64", ""}).getPluginByDevice(pluginName);
        // -----------------------------------------------------------------------------------------------------

[Maxim_S_Intel]

Try fp16 version of your model for the GPU this is Model Optimizer conversion option).

Also make sure to run sufficient number of iterations

[nikos1]

+1 for fp16 !  We are seeing 25% to 30% faster execution with fp16.

Also the HD630 is a GT2 GPU with no so many EUs.  If you are on an Atom or a dual core slower CPU with a GT4 then the situation is much better for GPU.

In any case GPU will result in a much more power efficient and better load balance than running on 100% MKLDNN code on CPU.

In general even if GPU and CPU path are similar in speed GPU should be preferred as it is a more green option :-)

Cheers,

Nikos