The Nvidia Turing GPU Architecture Deep Dive: Prelude to GeForce RTX

I was under the impression from the initial coverage that they were actually getting something like 1 to 2 rays per pixel per frame. Perhaps even less.

From there they were using some sort of smoothing and/or temporal anti-aliasing to gather the data from multiple frames to get decent quality out of it.

Or are you proposing how many raise they would NEED to be able to do real time full raytracing? By the time we have that ability I imagine everyone will just want the 4K version anyway and we will be behind again.

Either way what nVidia has shown looks great. It’s too bad I’ll have to wait years to be able to use it as a console gamer.

The thing that strikes me with the RTX announcement is a general point about how important identifying useful intermediate steps are to bringing about new paradigms.

Unless a technological breakthrough is just around the corner, or you have the resources to push it forward (Space Race / Manhattan Project), it’s better to spend your energy identifying useful intermediate steps that you can offer to the market to fund & bridge yourself to the new paradigm. By having funding all along the way, you can gain a significant advantage to those pursuing the new elegance directly. [1]

A few examples:

STREAMING. People used to go to video stores to rent movies. As the internet emerged we dreamed of a new, more elegant paradigm: streaming. No more driving to a store, no more physical copy or late fees or damages, etc. But it was the clever discovery of an intermediate step - to use the internet to rent DVDs via mail - that created the brand and customer base that established the market leader (Netflix). Once internet infrastructure caught up, the switch was seamless. Meanwhile, there were many people who pursued streaming directly, but failed because they didn’t take the intermediate step (Broadcast.com).

ELECTRIC CARS. Traditional cars have super complex drivetrains. As battery tech improved, we dreamed of a new, more elegant paradigm of electric vehicles that improved efficiency and eschewed most moving parts, transmissions, exhaust systems, etc. But there existed a valuable, infrastructure-free intermediate step to get there: hybrids. Ironically they were even more complex, but they employed many new techs that helped move electric cars forward. Toyota has hugely benefitted from being the discoverer of this intermediate step. Obviously we now have Tesla leading the vanguard, but in the context of global development, nobody can predict if an Elon will show up in your generation.

AUGMENTED REALITY. Our current physical reality is awash in information - street signs, road paint, branding, menus, maps, clocks, games, warnings, nutrition labels, interfaces, etc. These are often completely irrelevant to us at a given time, and certainly not personalized to our needs. We dream of the day we can render overlays on our eyes to deliver the personalized versions of these (as well as entirely new things), which would over time mean our physical reality would get simpler, cleaner, and less wasteful. To deliver this elegant solution requires a lot of breakthroughs in display technology that are years if not decades away. Bundling SLAM tech into smartphones (looking at you Apple) and pursuing incremental use cases is an intermediate step that can grow the market until the point where the displays are ready, at which point those who best pursue this are likely to be the market leader.

Ray tracing is now on the same course. It's been known for decades that it is a far more elegant paradigm to reason about and generate images (vis-a-vis rasterization), but its compute requirements are so high that there's been this chasm people haven't been able to cross to get to ray tracing. Nvidia has now provided a bridge between these two worlds, by allowing raytracing of parts of the rendering pipeline alongside rasterization. Subsequent generations will slowly swallow the remaining parts that rasterization performs today. Basically the RTX is the graphics card equivalent of a Prius, growing into a full electric.

The addition of ray-tracing cores in the RTX line was a pleasant surprise to me, not only because it speeds the development of ray-tracing hardware, but because it showed intermediate steps existed that I didn’t know about before. It showed me we weren’t stuck waiting indefinitely for a promise of an elegant future that always seems a decade away. Pretty exciting.

[1] What I mean by paradigm is not just incrementalism or an evolution of one product into another (like iPod -> iPhone), but of a wholly different way to solve a problem that is more elegant / higher abstraction than previous ways, but that require breakthroughs in enabling technologies to get there. Rockets -> Space Elevators (material science; elegance is in ease of transport), Retail -> Online Shopping (internet; elegance is in personalization + stay-at-home), Coal -> Solar (energy storage; elegance is in eco footprint, low entry point & simpler tech), Driving -> Autonomous Driving (ML/sensors; elegance is in time savings / one less thing to learn & simplification+density of road infra). This is admittedly a fuzzy definition, and perhaps these examples are not perfect.

Why are RT cores so different than normal shader cores? What instructions/memory fetch does a ray trace operation do that couldn't be implemented as an added instruction set on the shader cores to navigate the volume tree?

From the article, the best I can see is the following, but can't that be solved with microcode or as an extra rendering pipeline stage?

> In comparison, traversing the BVH in shaders would require thousands of instruction slots per ray cast, all for testing against bounding box intersections in the BVH

I ask, because having more slightly larger general purpose cores seems better for traditional rendering and raytracing than dedicating all that die space to pure single-purpose RT cores.

It is more like a texture unit than a shader core. Tree traversal is a pointer chasing problem, where the CPU/shader core executes a few instructions, then starts a memory load and then sits idle for tens or hundreds of clock cycles waiting for memory. Cache prefetching can help but is usually not a good fit for tree traversal where there is very little computation per node.

It is all about memory latency hiding and not really about computation.

Interesting thought, but I'm not convinced.

Constructing a robust BSP tree is very expensive, because you need exact rational arithmetic and the number of bits grows quite large. You wouldn't want to do that at runtime, so you could only use this for static geometry.

If most of your geometry is opaque, you can already achieve the same reduction in lighting complexity with a Z pre-pass.