They want their *own* CUDA situation. if you're doing FPGA stuff, the Xilinx hardware is good and Vivaldi is really the only way to use it. AFAIK the open-source fpga situation is pretty far behind it (please correct me if I'm wrong, I work in a place that uses Xilinx FPGAs)

I seriously think the CUDA situation was set up intentionally.

Jensen Huang (Nvidia CEO) is related to Lisa Su (AMD CEO).

A decade ago, I saw a demo of some AMD skunkworks GPU datacenter tech, that could execute CUDA natively on AMD/ATI graphics cards. Initially was half speed, but having the flexibility was crazy amazing. Created a big buzz in the big iron and educational markets.

Where'd it go? Buried. You cant even find articles about it. Its a few comments on edtech datacenters.

Now look at AMD's graphics line. Where's their ROCm LLM tooling? It's a fucking joke. Its like they're intentionally sinking it for her Nvidia uncle Huang. And Su takes the cheaper CPU market and offers better features than Intel.