AWS: No CUDA capable GPU installed. Forcing GPU=0

[alexander-rakhlin]

Good day,

Can we expect neon to run on AWS g2.2xlarge instance, GRID K520?

I have Ubuntu 14.10, CUDA 7.0, CUDA samples okay, and nvcc is okay too:
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Mon_Feb_16_22:59:02_CST_2015
Cuda compilation tools, release 7.0, V7.0.27

Still "No CUDA capable GPU installed. Forcing GPU=0", for each option
GPU=nervanagpu
GPU=cudanet

[scttl]

Hi,

AWS doesn't support newer Maxwell based GPU's, so the nervanagpu backend will not work, but cudanet should.

What was the exact syntax of the command you were trying to run? Did you try make -e GPU=cudanet install? Were you trying to run as sudo?

[alexander-rakhlin]

Hi,

I set in setup.cfg:
GPU=cudanet

And tried both:
sudo env “PATH=$PATH” env “LD_LIBRARY_PATH=$LD_LIBRARY_PATH” make install
and
sudo make install

I didn't try:
sudo make -e GPU=cudanet install
neither
make -e GPU=cudanet install

Should it work?

[scttl]

If you modified the GPU line of setup.cfg then that should perform the same functionality as make -e GPU=cudanet so I don't think the result will differ.

Even though nvcc --version appears to work when you run from the command line, it looks like the error you are reporting comes from the return value from that exact test. In the top-level Makefile, the line is:
ifneq ($(shell nvcc --version > /dev/null 2>&1; echo $$?), 0)

Out of curiosity what does running: $(nvcc --version > /dev/null 2>&1; echo $?) and sudo $(nvcc --version > /dev/null 2>&1; echo $?) from the command line report?

[alexander-rakhlin]

I have to wait for reasonable price to check it (something unusual goes on with AWS GPU spot instances recently, up to $6.5/hour for g2.2xlarge)

[nervanasys]

Tracking this under #80 going forward.