While running the command make pycaffe, I encountered the error below:

NVCC src/caffe/solvers/adadelta_solver.cu nvcc fatal   : Unsupported
gpu architecture 'compute_20' Makefile:594: recipe for target
'.build_release/cuda/src/caffe/solvers/adadelta_solver.o' failed make:
*** [.build_release/cuda/src/caffe/solvers/adadelta_solver.o] Error 1

System Information

OS: ubuntu: 16.10

CUDA 8.0

cuDNN: 6.0 


         -gencode arch=compute_20,code=sm_20 \
        -gencode arch=compute_20,code=sm_21 \
        -gencode arch=compute_30,code=sm_30 \
        -gencode arch=compute_35,code=sm_35 \
        -gencode arch=compute_50,code=sm_50 \
        -gencode arch=compute_52,code=sm_52 \
        -gencode arch=compute_60,code=sm_60 \
        -gencode arch=compute_61,code=sm_61 \
        -gencode arch=compute_61,code=compute_61

Can anyone help me?

  • small change cuDNN version is 8.0
    – Asha Datla
    Sep 29, 2017 at 6:27

3 Answers 3


As for me, I had to comment out the -gencode arch=compute_20 in Makefile.config:

CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
    -gencode arch=compute_35,code=sm_35 \
    -gencode arch=compute_50,code=sm_50

I stopped at 50 because CUDA's deviceQuery showed me Capability Major/Minor version number:

/usr/local/cuda/samples/bin/x86_64/linux/release/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "GeForce GTX 960M"
  CUDA Driver Version / Runtime Version          9.0 / 9.0
  CUDA Capability Major/Minor version number:    5.0
  Total amount of global memory:                 4044 MBytes (4240965632 bytes)
  ( 5) Multiprocessors, (128) CUDA Cores/MP:     640 CUDA Cores
  GPU Max Clock rate:                            1176 MHz (1.18 GHz)

Then compilation and tests went well.


I had the same issue this morning. After installing CUDA and cuDNN, restart was required (as suggested here https://groups.google.com/forum/#!topic/caffe-users/WDOD3E04Avg), so that CMake properly detected set variables. So just make sure CUDA and cuDNN are properly installed and restart your system. If you still get the error, you might have a GPU that only supports compute capability 2.0, so I guess you could try CUDA 8.0 which supports it. You can check your GPU here: https://developer.nvidia.com/cuda-gpus

I can confirm that the tests were run successfully on my PC with CUDA 9.0 and cuDNN 7.0.2 enabled. After restart, the GPU architecture was automatically set to sm_50. I have a GTX 750 Ti, which according to documentation supports CUDA 5.0. So the configuration seems correct now! Here is the command for testing:

make runtest

If you get any errors while compiling the tests, you may try:

make runtest clean

This example also worked for me and it's more than 7x faster (60 seconds) than with OpenBLAS with 8 CPU cores (450 seconds)!

  • Thanks a lot for your notice that CUDA Driver v8.0 supports compute capability 2.x. It will give a second life to life my old GPUs (9600 and 410M).
    – 18augst
    Aug 14, 2018 at 23:37

I also had this issue on my Jetson TX2, when installing NVcaffe (running make -j4).

The instructions in the nvidia jetson forum, here, say to replace:

-gencode arch=compute_61,code=sm_61


-gencode arch=compute_62,code=sm_62

in makefile.config. However, that line was not in my config file since, following the instructions, I pulled caffe-0.15, which doesn't contain that line. So in the end, what worked for me was replacing the following in my config file:

CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
        -gencode arch=compute_20,code=sm_21 \
        -gencode arch=compute_30,code=sm_30 \
        -gencode arch=compute_35,code=sm_35 \
        -gencode arch=compute_50,code=sm_50 \
        -gencode arch=compute_50,code=compute_50


CUDA_ARCH := -gencode arch=compute_50,code=sm_50 \
        -gencode arch=compute_52,code=sm_52 \
        -gencode arch=compute_60,code=sm_60 \
        -gencode arch=compute_62,code=sm_62 \
        -gencode arch=compute_61,code=compute_61

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.