Slashdot Mirror
NVIDIA's Ray Tracing Tech Will Soon Run On Older GTX Cards (engadget.com)
NVIDIA's older GeForce GTX 10-series cards will be getting the company's new ray-tracing tech in April. The technology, which is currently only available on its new RTX cards, "will work on GPUs from the 1060 and up, albeit with some serious caveats," reports Engadget. "Some games like Battlefield V will run just fine and deliver better visuals, but other games, like the freshly released Metro Exodus, will run at just 18 fps at 1440p -- obviously an unplayable frame-rate." From the report: What games you'll be able to play with ray-tracing tech (also known as DXR) on NVIDIA GTX cards depends entirely on how it's implemented. In Battlefield V, for instance, the tech is only used for things like reflections. On top of that, you can dial down the strength of the effect so that it consumes less computing horsepower. Metro Exodus, on the other hand, uses ray tracing to create highly realistic "global illumination" effects, simulating lighting from the real world. It's the first game that really showed the potential of RTX cards and actually generated some excitement about the tech. However, because it's so computationally intensive, GTX cards (which don't have the RTX tensor cores) will be effectively be too slow to run it.
NVIDIA explained that when it was first developing the next gen RTX tech, it found chips using Pascal tech would be "monster" sized and consume up to 650 watts. That's because the older cards lack both the integer cores and tensor cores found on the RTX cards. They get particularly stuck on ray-tracing, running about four times slower than the RTX cards on Metro Exodus. Since Metro Exodus is so heavily ray-traced, the RTX cards run it three times quicker than older GTX 10-series cards. However, that falls to two times for Shadow of the Tomb Raider, and 1.6 times for Battlefield V, because both of those games use ray tracing less. The latest GTX 1660 and 1660 Ti GPUs, which don't have RT but do have integer cores, will run ray-traced games moderately better than last-gen 10-series GPUs. NVIDIA also announced that Unity and Unreal Engine now support ray-tracing, allowing developers to implement the tech into their games. Developers can use NVIDIA's new set of tools called GameWorks RTX to achieve this.
"It includes the RTX Denoiser SDK that enables real-time ray-tracing through techniques that reduce the required ray count and number of samples per pixel," adds Engadget. "It will support ray-traced effects like area light shadows, glossy reflections, ambient occlusion and diffuse global illumination (the latter is used in Metro Exodus). Suffice to say, all of those things will make game look a lot prettier."
NVIDIA explained that when it was first developing the next gen RTX tech, it found chips using Pascal tech would be "monster" sized and consume up to 650 watts. That's because the older cards lack both the integer cores and tensor cores found on the RTX cards. They get particularly stuck on ray-tracing, running about four times slower than the RTX cards on Metro Exodus. Since Metro Exodus is so heavily ray-traced, the RTX cards run it three times quicker than older GTX 10-series cards. However, that falls to two times for Shadow of the Tomb Raider, and 1.6 times for Battlefield V, because both of those games use ray tracing less. The latest GTX 1660 and 1660 Ti GPUs, which don't have RT but do have integer cores, will run ray-traced games moderately better than last-gen 10-series GPUs. NVIDIA also announced that Unity and Unreal Engine now support ray-tracing, allowing developers to implement the tech into their games. Developers can use NVIDIA's new set of tools called GameWorks RTX to achieve this.
"It includes the RTX Denoiser SDK that enables real-time ray-tracing through techniques that reduce the required ray count and number of samples per pixel," adds Engadget. "It will support ray-traced effects like area light shadows, glossy reflections, ambient occlusion and diffuse global illumination (the latter is used in Metro Exodus). Suffice to say, all of those things will make game look a lot prettier."
What kind of admission is this!?
for my old GTX 240. It sucks. Some games wouldn't let you turn it off, and since there was no hardware acceleration it all ran on my CPU. I was running a GTX 240, you can bet my CPU couldn't do physx.
By all accounts Ray Tracing already cuts framerates in half. I can't imagine a world where this works.
Hi! I make Firefox Plug-ins. Check 'em out @ https://addons.mozilla.org/en-US/firefox/addon/youtube-mp3-podcaster/
Yet another reason to take a pass on 2080. And 2060 doesn't even need another reason.
Then there is this. Probably, Radeon VII is the ideal platform because memory bandwidth is everything.
When all you have is a hammer, every problem starts to look like a thumb.
> 18 fps at 1440p -- obviously an unplayable frame-rate.
Mother f**kers aint seen me overclocking my celeron to 233MHZ with a box fan to cool it to get 25 FPS in Quake.
You probably want to use a neural network for that.
You will need a large dataset, but you're on the freaking internet, so i bet it's not that hard to find it.
Then you can create a "thisbuttholedoesnotexist.com"
I played through Metro Exodus with a 2080 ti on a shiny new machine on a 4K monitor.
I remained unaware through the whole experience as to whether RTX was on and if it was, what difference it made.
Side by side you might be able to tell, but you don't play games side by side.
Maybe it looks a little better, but if it does, I don't care. The game was pretty good and fun to play, albeit with a stupid ending, which seems normal these days. Issues of frame rate, RTX, DLSS or anything else never really impinged on anything I care about.
I should use this sig to advertise my book ISBN-13 : 978-1501515132.
Surely nVidia's decision to enable their older cards to run ray tracing has nothing to do with Crysis demoing real time ray tracing on AMD GPUs a couple days ago. As Cryengine has shown, real time ray tracing can be done in software without the need for specialized hardware accelerators. This kinda makes the main selling point of the GeForce 20 series more or less moot.
Sure, it allows for super-realistic reflections and shadows, but we can fake those, and spend the resources on other effects or simply a higher resolution and better frame rate.
Comment removed based on user account deletion
real time ray tracing can be done in software without the need for specialized hardware accelerators
Let me rephrase that for you, {...} "transform and lighting can be done in software without the need for specialized hardware accelerators".
Actually only this last part is comparable to the discussion.
When T&L was introduced by Nvidia, it wasn't offering magnitudes more than what a beefy CPU with the latest SIMD extensions could offer.
Transforming a larger collection of geometry was equally possible by adding separate specialised single puprose blocks to the GPU, or adding even larger faster SIMD with multithreading capabilities to the CPU.
It's only later, when the shader got unified (it's not anymore extra specialized single purpose blocks, it's blocks that can be used for pixel shaders too, depending on the load) and at a time where most of the data stays resident on the card.
(In fact, if you squint at it - or just are used to CUDA - they ARE the apex evolution of xxx-large SIMD with epically wide multi-threading. Except they are "geographically" situated closer to the graphic card than the main x86 CPU)
Here we have again a somewhat similar situation.
Nvidia is adding separate different units to their GPUs, specialized to handle deep-neural-nets (different data types and operations than the general purpose shaders used for graphics), because a sizeable chunk of their market (the data center market) asks them to.
But how do you market the same GPUs to gamers ? Well, turns out you can run a specially trained denoising net to make graphics using fewer rays.
And now you have a selling point in conferences.
But it turns out, it's also possible to run the de noising using other type of approaches (whatever AMD is using) and not rely on "exclusive for neural net" separate cores.
So now, they are stuck, and try to push the technology, but running the denoising on the normal graphic shader (the same way most AI research has been done with GPGPU before "dedicated tensore cores" were a thing). Turns out, on older hardware it's a bit taxing on the overall computing power available for graphics (for obvious reasons), and depending on how much raytracing is goin, it slows down.
TL;DR: So it's not a question of "before high speed dedicated hardware could do it, it was also possible with software on CPU (although slower, but nobody noticed it yet, because no game leveraged that many yet)" like rendering textured triangles.
It's a question of "at the current level, there's more than a way to do it. There are different hardware solution to this problem, not all of them necessitating new single-purpose cores that can only do that (x)", like whether a new type of single-purpose shader taking precious silicon real-estate on the GPU, or new mass-computation extensions on the CPU should be the one handling T&L.(*)(^).
----
(*): and in the end, it turned out to be "in between". A new class of "more general purpose mass-computation blocks" but "on the GPU" (shared shaders) that could also be repurposed for even different needs (the rise of the GPGPU... all the way to modern AI on GPUs and even Raytracing on standard shaders".
(x): it's single purpose from the point of view of gamers.
That's why it's fast on RTX cards: it's using an otherwise left untouched part of the GPU and not taxing the normal graphical shaders.
Of course those cores can also run neural-net in data centers (so a different purpose) which was the whole reason for Nvidia adding them in the first palce.
(^): to take your Commodore 64 metaphor, the idea isn't *hardware* sprite (C64) vs. *software* on CPU (PC).
it's hardware *sprites* vs. hardware *blitter* vs. hardware *affine-textured stripe*.
with different advantage brought by each solution.
(here: the second one is more complex, but gives you more flexibility.
the last one is even more complex, but gives you a roto-zoomer for free, and could even be used to implement primitive polygons)
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
I am not dissing it; I am just trying to understand its relevance - not being a gamer, I guess that it won't be very relevant to me. Nevertheless, I would be interested to learn in what other areas this is likely to have an impact.
... or three times as quick?
1.6, 2 and 3 times as fast or 1.6, 2 and 3 times faster?