Inference time for EfficientFormerV2 on 2080Ti

[zcbfpramtqs55675]

Hi, I tested EfficientFormerV2-s0 and EfficientFormerV2-s2 on 2080Ti, the input size is 1x3x224x224, and got the result as follows:
EfficientFormerV2-s2: about 24ms/per input,
EfficientFormerV2-s0: about 22ms/per input
Is this reasonable? seems different from A100 results in your paper. Any reply is appreciated, thx.

[giantmonkeyTC]

I tested V2-s1 on 3090 and got 14ms/per sample.
I don't know why either. If you figure it out, could you please reply to this thread?

[alanspike]

Hi, we use TensorRT to benchmark the latency. Here is the docker image.

[caixiongjiang]

I used GeForce RTX3060 on my segmentation model using the EfficientFormerV2-S0 backbone and the PoolFormer-S12 backbone. The FPS results are 61 frame/s and 108 frame/s. I don't think this backbone is very universal in calculating the time, and it is specially designed for iPhone. It is similar to such segmentation models as ENet, although the number of parameters and the amount of computation are relatively small, the actual computation time is very long, which may be related to the fact that Pytorch has no acceleration for this part of the computation.