Technology

Software like wgpu makes it much easier to close the gap between various GPUs. New compute languages that are backend-agnostic are appearing, in the same vein as taichi-lang, that make it significantly easier to make high-performance gpu kernels deployable anywhere.

The compute groundwork for crossplatform tensor calculations is already here. Inference is already doable on any device. Training is not far behind. As a side-effect of this, processing on the GPU in every capacity, like physics, novel rendering techniques, or whatever else the imagination can muster, is now within grasp of "average" programmers.

If you have always been intimidated by GPU programming, I urge you to take another look now. The landscape is radically different. The software moat everyone talks about with NVIDIA is smoke-and-mirrors. Cuda is old news, though I am speaking to the actual code landscape here, not the common mental consensus.

What we lack now is cheap video cards that have high memory. I believe the current cards are overpriced by about 10 - 100x what they should be, because this profit situation is extremely temporary. Just as pens were once thousands of dollars, these compute devices will be collapsing in price.

I welcome China building cheaper video cards. Hopefully we will all benefit from it before any robot wars break out.