What Happens to Rasterization in the Era of Ray Tracing?

[Color me Dan]

It's been roughly 6 years since the introduction of the first consumer cards with dedicated ray tracing hardware, and all of these years later we're finally seeing games that require it to work. I imagine that in another 6 years time we'll see massive gains to the ray tracing hardware, and increasingly modest development towards raster performance. So my questions are basically:

A) Will hardware and software reach "peak raster" in the near future? The point where progress basically halts for raster before silicon is increasingly dedicated strictly to ray tracing?

And B) what will happen to all the games produced that require raster hardware? Will future ray traced focused hardware retain, or be capable of emulating in software, features necessary to play them?

 

[pcchen]

I think it'd be interesting to see whether GPU will be less focused on rasterization hardwares, and more on computation hardwares.
Some "rasterization" hardwares will probably always be there, such as triangle setup, depth buffer, and maybe textures. There might be case for replacing texture hardwares with computation units although it probably won't be very soon.
However, as one can easily see that today very few GPU are actually limited by those traditional "rasterization" hardwares such as triangle setup and texture units, so there will be much less incentive to improve on these fronts. For example, 4090's theoretical pixel fillrate is more than 450Gpix/s, and that's enough to fill 4K at more than 50,000 FPS. That's obviously way more than enough for practially any situation.

 

[DavidGraham]

We are not in the era of pure ray tracing yet, we are doing hybrid raster + ray tracing. Even path traced games are still hybrid, this isn't going to change anytime soon. Ray Tracing still relies on general purpose compute hardware, and I think it's going to remain this way for a long time.

Judging by the current trends we are more likely to head in a direction of more hybrid rendering involving raster + ray tracing + machine learning with variable degrees. I think machine learning hardware could see more rapid adoption than the other two.

 

[trinibwoy]

I wouldn't count texture mappers as raster hardware in this context as their function is equally applicable to RT. This is just a guess but I think with RT there will be less need to write out pixel buffers to memory. Shadow maps and cube maps will be gone. Gbuffers and depth buffers will likely be useful for a very long time for primary visiblity and various post processing effects. Hardware raster, tessellation and blending (ROPs) seem like good candidates for being replaced by software in the medium term if and when there is less demand for classic pixel rasterization in games.