What can I do with RTX 5090 that I couldn't do with RTX 4090

I'm no expert here. But the 5090 has 33% more TFLOPS, so if the whole model fits in VRAM then once you have the model loaded, you'd expect compute to be about 33% faster. That's going to give you roughly 33% increase in tps when inferencing, or so I understand it. Training is a different beast and will have a different trade-off. When you say "80% higher bandwidth" what exactly are you referring to? PCIe bandwidth? For just using a trained model, that's almost irrelevant, because you load the model parameters once and don't really care how long it takes (within reason). What matters there is raw ability to crunch data, so long as it all fits into VRAM (as soon as it doesn't all fit into VRAM it's a different story, of course, as you have to shuffle data back and forth between VRAM and main memory).