~~ * Information as of October 24, 2020. ~~ ** PyTorch 1.7 was released on October 28th, and it is compatible with CUDA 11 and can be used with RTX3090. ** **
Therefore, if you simply upgrade to PyTorch 1.7, you can use it with RTX3080 / 3090.
The following is old information.
PyTorch 1.6 does not currently support RTX 3080/3090 and throws the following error.
GeForce RTX 3090 with CUDA capability sm_86 is not compatible with the current PyTorch installation.
The workaround is to install the development version of PyTorch 1.8.
pip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html
Although it is cu110, there is no problem because it works even if you are using cuda11.1.
NVIDIA-SMI 455.23.05 Driver Version: 455.23.05 CUDA Version: 11.1/Ubuntu 20.04
I used to use 2070 SUPER, but the learning speed is about 3 times faster. Also, jax worked normally.