The Perils of Ray-Tracing Performance Marketing and Peak Ray Flow


Ray tracing is arguably the most significant change to the real-time graphics pipeline in the last decade. It gives developers new tools in the box of tricks they use to carefully persuade today’s GPUs to draw immersive, realistic, high-quality scenes at high performance.

Within the last couple of years, products have hit the high-end PC market that support Microsoft’s DirectX Ray Tracing (DXR) specification. A number of high-profile games that use DXR showcase the potential of the technology.

The great folks at Khronos have also recently released their view of ray tracing as it applies to Vulkan. The resulting API will feel familiar to students of DXR, giving developers another ray-tracing API to target and a means to expand over time to platforms that aren’t Windows.

However, at the time of writing, all of the public implementations come from a single vendor, Nvidia. Consequently, developers have a rather narrow view of what makes an accelerated ray-tracing implementation based on DXR or Vulkan ray tracing tick.

That’s no fault of Nvidia, the graphics giant taking advantage of their ability to strong arm their view of ray tracing into DirectX and therefore robustly influence the resulting capabilities in Vulkan. Any other vendor in Nvidia’s position would do the same, and the company’s ability to productize the output of its incredible research team is second to none.

However, if that’s how a new graphics capability comes into being, driven almost completely by a single vendor in the beginning, resulting in a narrow spectrum of implementations with limited market share, developers need to be careful as that landscape of implementations broadens over time.

Multiple Design Paths

That’s especially true when it comes to a largely black box feature like ray tracing. What do we mean by black box? For any given hardware and software specification in the semiconductor world, and that isn’t limited to GPUs, there’s a huge design space you can explore to implement it.

There’s no one way to build a CPU for example, leading to a wide range of practical instruction set architectures (ISAs) that can all run roughly the same software. GPUs enjoy an even wider range of potential implementations from different vendors because there’s no standard ISA from which the industry designs its implementations.

GPU vendors take advantage of that to innovate more rapidly under the hood compared to CPUs. That’s because the combination of hardware and software at work can more readily hide what’s under the hood doing the work at the silicon level.

So, when Nvidia proposed what became DXR to Microsoft, when specifying the programming model and API that developers need to target to perform ray tracing in their renderer, they left certain parts of it almost completely undefined.

That lack of guidance in parts of the specification is completely on purpose. Why? It allows the underlying implementation of that bit of the specification to be entirely specific to a given GPU and its driver. In fact, parts of DXR and Vulkan ray tracing don’t even need to run on the GPU at all!

And therein lies the rub: because of that, unlike almost every other feature ever added to the real-time graphics pipeline in GPU history, ray tracing will be very difficult for developers to target in a uniform way.

That’s further complicated by the fact that ray tracing is also bifurcating from the developer’s point of view, too, with a simpler model having emerged from the full ray pipelines first specified. This allows for easier implementation of some specific effects, particularly ray-traced shadows.

Advice for Developers

So, with two ray-tracing programming models for developers to target and still only one generation of shipping accelerators to try it out on, and with key parts of DXR and Vulkan ray tracing specified as completely implementation-dependent, what should developers be aware of as things develop in the ray-tracing acceleration space?

As a company developing ray-tracing acceleration for its future GPUs, where our solution will have microarchitectural performance characteristics that are necessarily different—and better by far, we hope—to those implementations from other vendors, we have some simple advice.

First, avoid any marketing claims about ray-tracing performance since there’s no standard way to describe it. The black box nature of the ray-tracing specifications makes it almost impossible to do so anyway. Let’s pick on one of the most impressive GPUs ever made to push home the point: Nvidia’s TU102 in GeForce RTX 2080 Ti form. It’s a competing product so I shouldn’t really say this, but woof, it’s one heck of a GPU.

Nvidia says there’s a peak ray flow of 10 Grays/sec in that product, which is an incredibly impressive headline number. Sadly, though, it doesn’t actually say anything concrete that you can understand as a consumer, developer, or anyone else interested in the technology and how it works.

Does it only measure miss-only rate with empty rays, or is it measuring real-world usable ray flow on rays with a proper payload? It’s a marketing number, so it’s definitely a lot closer to the former than the latter, unfortunately. It also doesn’t tell you anything about the location of the performance cliffs. Nor does it tell you how ray-tracing throughput degrades with complexity of ray, hierarchy, or any of the other moving parts of the ray-tracing system, and how it’s all integrated with the rest of the rendering workload.

As a result, you’re left with testing and measuring as your only realistic way to determine what’s possible with a given level of ray-tracing performance in a GPU. That’s because Nvidia’s 10 Grays/sec won’t mean the same thing as the same number quoted by Imagination, or any of the other vendors that will try their hand at useful DXR or Vulkan ray-tracing acceleration for that matter. It won’t even mean the same thing generationally for Nvidia!

Lastly, does that mean the graphics market now has a need for public software that tests meaningful peak microarchitectural ray flow to expose implementation-specific details, along with some measures of real-world performance in idiomatic game-like settings? Absolutely, and we’d love to talk to anyone who wants to tackle that problem, since it’s a big missing piece of the puzzle.

Rys Sommefeldt is Senior Director of Product Management at Imagination Technologies.



Source link