Comment by kouteiheika

> I firmly believe that Nvidia doesn't want the general public to ever have better hardware than what is current as people could just run their own local models and take away from the ridiculous money they're making from data centers.

You're underestimating them. They don't even want rich professional users to own hardware that could compete with their datacenter cash cow.

Take RTX 6000 Pro, a $10k USD GPU. They say in their marketing materials that these have fifth-generation tensor cores. This is a lie, as you can't really use any 5th-gen specific features.

Take a look at their PTX docs[1]. The RTX 6000 Pro is sm_120 in that table, while their datacenter GPUs are sm_100/sm110. See the 'tcgen05' instructions in the table? It's called 'tcgen05' because it stands for "Tensor Core GEN 05". And they're all unsupported on sm_120.

[1] - https://docs.nvidia.com/cuda/parallel-thread-execution/#rele...