Even a relatively accessible DGX Spark shows this kind of computing power. A 4090 would probably be twice as fast... Work that used to take hours on a Mac Pro can now be done in tens of minutes with CUDA. (The data is about ten years worth of transaction records and hundreds of institutional evaluation data...)
CUDA is amazing. But it's frustrating that even with this processing speed, LLMs can only achieve basic performance.
For AGI to arrive, we probably need something like an RTX PRO 12000 instead of an RTX PRO 6000. I have no idea how much processing power will be needed.
