A lengthy thread I wrote just to write. Feel free to ignore.

1/ The future of consumer GPUs is probably less ASICy than people think. I think we're nearing the end of growing non-general purpose compute/cache die utilization in consumer GPUs. Here's why:

2/ Real world RT core usage is far from peak theoretical perf. On Nvidia, they're cache starved. The way to improve performance for Nvidia is to add more cache to the die a la infinity cache. On AMD, it's just to improve their RT core design, which isn't great.

More importantly,

3/ real world RT core usage is very naive. It's often collect n sample per pixel with m reflections w/irradiance caching, & then run A-SVGF to denoise. This usually works well, but sucks w/moving objects, fine details like leaves on distant trees, and the combo of leaves in wind.

4/ But research already shows that you can get pretty drastic improvements w/adaptive sampling in that case.

Beyond that, adaptive sampling research usually focuses on general cases. But consumer GPUs use RT for e.g. games, Blender, etc. These aren't super general. The engines

5/ underpinning these applications have access to a lot of information. For example, developers can add hints to objects to suggest to the renderer that they be sampled more or less.

Better exploitation of temporal data and ray tracing statistics can make a lot of the rays cast