Confused: TVM vs TF, performance compare


My cuda info:

cuda: 9.0
cublas: 9.0
cudnn: 7.0

I run resnet50(2242243) using TVM and TF at the same computer, time result is as follows:

  1. TVM: 3.5ms
  2. TF(1.9.0): 2ms

I wonder to know why TF is faster than TVM so much?


Can you share more details about your TVM setup (e.g., are you using a cuda target or cuda -libs=cudnn, the model of the GPU, etc…)? You may be able to improve the performance by tuning for your specific GPU.



VGA compatible controller: NVIDIA Corporation GM200 [GeForce GTX TITAN X] (rev a1)

when compiling TVM, I open USE_CUDA and USE_CUDNN switch.

can you help me to analyze the performance reason?


Does print the cuDNN profiling results the first time you run inference (just to verify that the JIT compilation to cuDNN is happening)?