GPU Frameworks Inspection

Here are investigated a few properties of the current GPGPU programming frameworks, and the landscape in general.

Initial

The original plan was to see if there is any framework:

1. as platform generic as possible (not NVIDIA specific)
2. with C/C++ API
3. allowing to manage memory manually
   • to support performance critical and non-standard use case
4. if possible, with a Python API out of the box

The outcome is that there is no such a framework, but many options each one with some limitations.

In particular, there are many good and interoperable Python abstractions:

• CuPy
• TensorFlow
• PyTorch
• Numba
• ...

And a couple of standards for tensor exchange between frameworks (Numba's __cuda_array_interface__ and DLPack).

For more details see the dedicated note.