Also, upscaling doesn't need ML hardware. It's fancy to call everything ML and AI but DLSS actually uses the dedicated HW (Tensor cores) just for a ONE very small step in the whole upscaling process.
It can run perfectly fine on the CU's. DLSS isn't better than FSR because of the ML hardware, Nvidia had simply more time to work on it and/or thrown more resources at the problem so far.
ML hardware would be a waste of silicon in a gaming console.
AMD's next-generation graphics pipeline (NGGP) model was VEGA and it was carried over to NAVI10.
VEGA's NGGP is broken. AMD's NGGP is a failure.
NVIDIA's RTX next-generation graphics pipeline model
just works. AMD has to follow NVIDIA's RTX innovation but with incomplete RDNA 2 RT core implementation.
NVIDIA's Ray Reconstruction (included as part of DLSS 3.5) offers a suite of AI rendering technologies powered by Tensor Cores on GeForce RTX GPU. The current denoise pass for RT is done of shaders.
You're wrong with the "ML hardware would be a waste of silicon in a gaming console".
Your argument will change when AMD reaches feature parity with NVIDIA's RTX.
A Phoenix APU has an AVX-512 Vector Neural Network Instructions (VNNI) extension from Zen 4's AVX-512 implementation, discrete AI accelerators for RDNA 3, and a discrete NPU (Neural Processing Unit).
Future PC APUs scales from the Phoenix APU baseline. AMD's weakness is software while the hardware has good bang per buck value.
For RDNA 3's AI accelerator hardware, AMD is missing the software stack for AI-driven Ray Reconstruction and DLSS equivalent.