|
Decided to split the off-topic part off into a reply so that it didn't distract from the answer! In terms of over-optimisation forcing a certain technologies to be developed and others to be ignored, one example I'm very familiar with is computer graphics. I'd written a TON of stuff here, but decided to simplify it as it was labouring to specific a point. But our computer graphics state-of-the-art was roughly along these lines: drawing all edges of polygons, hidden-line removal (Sutherland), clipping intersecting polygons (Hodgman), filling polygons with a single colour, *, Gouraud shading, Gouraud shading with smaller triangles, Phong shading with bigger triangles, texturing, fixed texture and lighting pipelines, pixel shaders, vertex shaders. I'll also add compute shaders too, but that was more of a generalisation of what people were starting to do with pixel shaders operating on data that wasn't really pixel data. Now, you'll notice my * around the time of single colour filled polygons... this might not be the correct place to put the *, but around this point some people started experimenting with ray-tracing and got amazing results, just incredibly slowly. These were seen as the "gold standard", but because drawing triangles was much faster, this is where the money continued to be poured into, optimising and optimising this special case, discovering more techniques to "approximate" the right image, but trying to avoid the hard work of actually rendering it. Over time, things have got closer and closer to ray tracing, except transparent and shiny objects have always been the achilles heel. Fortunately industry's interest in ray-tracing has resumed, and now compute shaders are general enough that they can be used, but they're still orders of magnitude slower because the renderer needs to consider the entire scene not just a triangle at a time, so you need to store the scene in some kind of tree that's paged in on demand and different latencies for different pixels causes problems for the SIMD architectures. We're starting to see more and more consumer-level hardware with decent ray-tracing performance now, but it's been a decade of lost time in terms of optimisation from where it could have been if the entire market hadn't been competing only in making triangles rasterise more quickly. In the ray-tracing space, we still see that it's too slow to create perfect images (for very complicated scenes with lots of shiny surfaces and few lights, you might need thousands of rays per pixel to just get a handful that actually reach a light source), so we've invented all sorts of approaches to cover it up - whether it's training an ML model to guess the real colour for black pixels from neighbouring ones, or re-projecting pixels from a previous frame to fill in the games, etc. Personally, I can't help but think the real breakthrough in performant raytracing will come from tracing light from the light sources instead. This wasn't done traditionally because potentially it's even more expensive than tracing backwards from the pixel, but should be more accurate when there are multiple light sources. But even the latest batch of hardware is all focused on raytracing, which I think is missing the biggest trick of all - they could be using cone-tracing as a first approximation and then subdividing the cone into smaller and smaller chunks until they're approximately pixel sized. None of this is new, it's just not what the larger industry is doing right now, because it's cheaper and easier for them to do rays instead. |