It's a very blurry thing to answer as the other replies highlight. If you accept data points from the 90s, the work that Kaze Emanuar is doing on Nintendo 64 e

It's a very blurry thing to answer as the other replies highlight. If you accept data points from the 90s, the work that Kaze Emanuar is doing on Nintendo 64 engineering improvements is a good indicationthat it's possible. Improved compilers, more modern algorithms, rethinking rasterization and how everything happens on device. There's a lot of "what if we had infinite time to optimize for this stack" for sure. His comparison of various approaches to solve fast inverse squaee root on N64 hardware is really impressive.