The gap between theoretical computing power and real computing power？

[liu0527aa]

We use the code to test the performance in NVIDIA NX. However, we found that there is a gap between the calculated real computing power and the theoretical computing power of the device.

Here is the result(we take Resnet50 as an example. ): it requires 8.2G FLOPs. According to the nvidia. According to the introduction of nvidia nx, we can find its computing power is 21 TOPS(int 8). So, we can get 21 *1000/8.2 =2560 FPS (with the GPU and 2 DLA). However, in the actual test, we found that the real performance is not so good.

First, we only use GPU to test. we get the best result FPS =894(batch size is 894) using GPU +2DLA we get FPS =1033（GPU BS=32，DLA BS=64）

Could you tell me the reason about the result?Thanks