1. Divide the world
NVIDIA has very successfully combined the PhysX physics engine with CUDA technology and Geforce graphics card to give birth to nVIDIA PhysX physics acceleration technology.
In order to compete with it, AMD launched the "Open Physics Project" based on AMD APP technology to contend with it. The Open Physics Project combines the three physics engines of Havok, Bullet and Pixelux DMM. Based on the OpenCL standard and AMD APP technology, the powerful parallel computing capabilities of AMD GPUs are used in games to accelerate physics calculations. At the same time, the Open Physics Plan also implements the Fusion concept. The CPU+GPU jointly performs physical calculations. The GPU is responsible for flexible material simulation, fluid simulation, explosion simulation and other large-scale physical simulation calculations.
2. Choose only one
If the graphics card is NVIDIA, pytorch needs to install the NVIDIA driver if you want to use the graphics card, and then download the corresponding according to the NVIDIA website: cuda (CUDA Toolkit 11.1 Update 1 Downloads) cudnn (NVIDIA cuDNN) needs to be registered and then installed according to the installation guide. Next, install pytorch. See Pytorth Introduction, Installation .
My graphics card is AMD, so I have to talk to CUDA byebye
The full English name of ROCm is Radeon Open Compute platform, and the goal is to build an ecosystem that can replace CUDA. The biggest difference between ROCm and CUDA is its openness. ROCm hopes to run on a variety of different hardware, and ROCm is completely open source.
The ROCm ecosystem supports a variety of open source technologies, including frameworks (Tensorflow / PyTorch), libraries (MIOpen / Blas / RCCL), programming models (HIP), interconnection (OCD) and Linux® Kernel upstream support.