Our software supports GPUs with the following compute capabilities:
- Pascal;
sm_60,sm_61 - Volta;
sm_70 - Turing;
sm_75 - Ampere;
sm_80,sm_86
The toolchain combination is non-exhaustible. We listed our experienced testbed below. Number in GPU-(GCC,CUDA) combination denotes major version of compiler ever tested.
| setup | arch. | SM | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GCC | 7.x | 7.x | 7.x | 7.x | 7.x | 7.x | 7.x | 7.x | |||||
| 8.x | 8.x | 8.x | 8.x | 8.x | 8.x | 8.x | |||||||
| 9.x | 9.x | 9.x | 9.x | 9.x | 9.x | 9.x | |||||||
| CUDA | 9.2 | 10.0 | 10.1 | 10.2 | 11.0 | 11.1 | 11.2 | 11.3 | 11.4 | 11.5 | 11.6 | ||
| P2000M | Pascal | 61 | 7 | ||||||||||
| RTX 2060S | Turing | 75 | 7 | 7 | 7 | 7 | 9 | 9 | 9 | 9 | 9 | 9 | 9 |
| RTX 5000 | Turing | 75 | 7/8 | ||||||||||
| RTX 8000 | Turing | 75 | 7 | ||||||||||
| RTX 3080 | Ampere | 86 | 7 | 9 | 9 | ||||||||
| P100 | Pascal | 60 | 7 | ||||||||||
| V100 | Volta | 70 | 7 | 7/8 | 7 | ||||||||
| A100 | Ampere | 80 | 7-9 | 7-9 | 7-9 | 9 | 9 | 9 |
It comes with GCC 9.3 and is confirmed workin giwth any CUDA 11.x: no extra setup is needed.
- CUDA 11.6 + GCC 11.2 confirmed not working (as of 2022 Feburuary).
- Because of
<functional>header inclusion error, even though it is officially supported. - Others experienced the similar issue: case 1, case 2
Fedora only provides single-version GCC, i.e., Fedora 35 + GCC 11.2. A workaround is use 3rd-party pacakge manager to install other version of GCC, e.g., linuxbrew, spack. A linuxbrew setup is given below
# defult path to install from .run file
export CUDA_ROOT=/usr/local/cuda
# install gcc
brew install gcc@9
# get the absolute paths
sudo ln -s /home/linuxbrew/.linuxbrew/bin/gcc-9 $CUDA_ROOT/bin/gcc
sudo ln -s /home/linuxbrew/.linuxbrew/bin/g++-9 $CUDA_ROOT/bin/g++
# consider to put these in ~/.zshrc or ~/.bashrc
export CC=gcc-9
export CXX=g++-9