If you really want to learn from first principles, I'd recommend writing a software renderer without any graphics API. If you're looking for a course, see [1]. Next up, extend the software renderer to render a 3D character model (an obj file will suffice) which [1] covers and animate that 3D model with skeletal animation all the while on the CPU. Use quake's md5mesh and md5anim files to render the skeletal animation from [2]. This just lays down a clear foundation of what a rendering pipeline looks like.
Then pick up something like Modern OpenGL and you'll easily intuit if you've put in the hard yards above, which parts of this pipeline are fixed and which are programmable on a hardware designed specifically to do 3D. It just so happens that the same is reusable for compute and there you go with all the D/ML jazz. After this, read up on the HW architecture of GPU's (the literature is sparse as these are vendor locked) but realtimhrendering books has chapters on it and you can find out why Vulkan, DX12, Metal are in vogue now and OpenGL has been deprecated.
[1] https://pikuma.com/courses/learn-3d-computer-graphics-progra...