Nvfortran installation

Hi, help me to start, please. I’d like to estimate, it’s worth to pass my codes to GPU. I’ve taken GeForce GTX 650, and install Debian 10.7, hpc sdk 20.11 and driver with NVIDIA-Linux-x86_64-460.32.03.run. But if I use option -cuda, I get an error message:
avp@D:~/work/NV/chapter1/increment$ nvfortran increment.cuf
NVFORTRAN-S-0155-Invalid value after compiler -cudacap flag 30
NVFORTRAN-S-1001-All selected compute capabilities were disabled (see -Minfo) (increment.cuf: 1)
0 inform, 0 warnings, 1 severes, 0 fatal for

I gess, there is some simple mistake, but which? I’ve spend a few of days, but I don’t guess.

The problem here is that compute capability 3.0, which is the architecture your GTX 650 uses, is no longer supported by NVIDIA. As such the NVHPC compilers no longer are able to generate code for this device. I’ll put in a request to see if we can get a better error message here so the issue becomes more obvious.

While you wont be able to run the binary unless you can get access to a newer device (the GTX 650 is over 8 years old and five generations older than our current products), you can compile the binary if you target a newer device via the flag “-gpu=cc35” (or one of the other support targets, see “nvfortran -help -gpu” for a complete list).

-Mat

Thank you. I assumed this might be the case, but I was not sure enough to spend money on a new card just for the experiment.

One option is to rent a system on AWS using the NVHPC Docker container from our NGC site: NVIDIA HPC SDK | NVIDIA NGC

I haven’t use AWS myself for a few years so don’t the exact cost but don’t think it’s too expensive. You can then experiment with your program to see if you think it’s them worth the investment of purchasing a new card.