Hi,
we use the tensorRT by onnxRT. This has dependencies to cudnn and cuda toolkit.
We extracted the really needed DLLs and found that theses DLLs together need about 2.5GByte (without onnxRT) - this is a problem for us and our customer for deployment.
So the question is, why some of the DLLs are so huge and if this is really needed or something is wrong. e.g.
cudnn_cnn_infer64_8.DLL ~574 MByte
cublasLt64_11.DLL ~531 MBytes
nvinfer.DLL ~441 MBytes
Best regards,
Oliver
The math binaries, in this case cublasLt, include many optimized kernels for different sizes, data types, and architectures.
1 Like