TensortRT vs NVCaffe

What are the trade-off between TensorRT vs NVCaffe?
Is there benchmarks comparing TensorRT and NVCaffe?


TensorRT has better performance than Caffe, but not all Caffe layer is supported.
Users need to implement non-support layer by them self.

For supported layer information, please check here:
[url]NVIDIA Documentation Center | NVIDIA Developer

We have some benchmark result for TensorRT1.0 v.s. Caffe.

But currently, we already have lots of improvement from TensorRT1.0:

  1. Faster device(TX2):

  2. TensorRT 2.1(together with cuDNNv6):