i would love to, but it's hard to discuss that which we know so little about.
Next gen lighting technologies - voxelised, traced, and everything else *spawn*
- Thread starter Scott_Arm
- Start date
I mentioned it repeatedly*, but with so little coverage and so little input from those who know, and zero experience in the field, there's very little to be said! I think we'd all love to hear from those who know about it - what was PVR's solution, how did it perform, and what was learnt from it. There's probably a lot of insider knowledge that's NDA'd though.
* eg September 27th
DavidGraham
Veteran
Thank you for voicing what many here feels already.
Last edited:
They did try. It wasn't fast enough. Those dabbling in compute-based volumetric lighting solutions found methods that worked and looked good, but were unusable for realtime applications. The mathematical theories existed but the hardware couldn't do anything with them. It's akin to using machine learning, your new delight! Machine learning and neural nets are nothing new, but now that we have hardware fast enough to implement them across a whole range of applications, people are exploring them again. Neural nets weren't used to develop image upscaling in 1982 despite the theories existing, because it wasn't practically possible. Now it is, and we have an explosion of research into new fields machine learning can achieve, because the hardware now enables it, and not because the theories have suddenly been discovered/invented. If the hardware didn't exist to run these AI techniques a thousand-fold faster than previously, the research into them woudln't be happening. And now that GPUs are 30x faster than the best possible in 2007, devs and experimenters can revisit old ideas and explore new derivatives.
That's a partial interpretation of what's happening in the industry.
ML algorithms didn't explode because of computational power, computational power isn't even 1/2 the equation. It's exploding because no one ever predicted that NN algorithms could scale as well as they do with more and more training data. Most algorithms have a peak performance that it hits eventually, and we've found that neural networks just keep getting better the more data we feed it. The computation of that data helps speed things up because we're feeding it so much, but same algorithm with significantly less samples would be a significant poor performer in image recreation vs say our standard algorithmic checker boarding. That being said, the storage of said data and how we are able to collect so much more data than before has been a significantly larger impact to ML than computational power. AI performance is based on it's samples, some are better with less, and some are better with more more. AI performance is not based on how fast the model is computed or executed except in real time applications - thus lets revisit the concept of FF acceleration - enter tensor cores.
Training speed is the factor that we're interested in when it comes to computational performance. The faster it can chew through more data that faster we can revise our models and try to extract more accuracy from them. Computation of neural networks is equivalent to compile time of a video game. Running a neural network takes significantly less resources by comparison.
ML is still very well supported by CPUs as the GPU industry continues to grow, I have to do my work on a corei7 laptop. You can just jump onto Azure ML and see how much of it is done on CPU clusters.
If not processing power, than data (storage and transfer rates). Either way, it's enabled by advances in technology, and not development of new theories. If not, why is ML exploding now rather than 20 years ago? "Because all this time on PC we've had the ability to use machine learning for a myriad of opportunities. No one tried it through Z80, 68000, x86, Pentium, Cell, x64, GPU compute. No one tried this." The theories had to wait for the tech to enable their exploration, no?
This makes sense in case the algorithm is solid enough, or there is a temporal requirement justifying FF.
Neither is the case with classic raytracing for games. But that's my personal opinion, coming from work at alternatives, and from knowledge there is no 'one and only' raytracing algorithm fast in any case. Driver does not know if i do coherent or incoherent rays, i do. And this really is just a simplified example. (sure, driver can guess, or receive hints etc. but the argument would hold if go go into details deep enough.)
There is Minecraft showing infinite bounces at a quality never seen before in realtime. There is Claybook with very fast RT. There is path traced Q2 running buggy but at 60fps on my non RTX GPU.
There is ancient stuff like this, running on prev gen console:
You are wrong. There always was research on realtime GI, which is the final goal.
Yes, sharp reflections as in BFV are something new. But the game does not look less wrong because of this addition. And that's the problem. I'm not impressed about performance either, and i see alternatives. Children Of Tomorrow is still more impressive for me.
Q2 is great, but i can do the same on PS4 hardware, even with infinite bounces and better handling of materials. That's why i personally consider RTX a wrong investment..
My work is not magic, and i took inspiration from the works of those researchers, including many from NVs circle. I was in contact with some of them - nice guys and helpful. I have no PHD, but i consider myself as a researcher as well, kind of.
As i am unable to proof my magic crap yet, we should not further argue. You win in any case so there is no need for frustration on your side. HW RT has landed and it will succeed. Even i will follow and help with that.
The real frustration is: It lasts 5-10 years until developers get what they need and request. See Andrew Lauritzens Talk about current state of compute in games. Quote: "Most of the problems I will discuss here were known 7 years ago"
https://de.slideshare.net/DICEStudio/future-directions-for-computeforgraphics
You just have to accept critique from devs - they will always rant about potential improvements and request for changes. At least i hope so.
I think its pricing thats the problem, RTX gpus are too expensive right now. Most cant afford a 2080, and therefore are perhaps annoyed by the fact nvidia raises prices their new tech. The lack of competition doesnt help it either.
Theres also the userbsse split, PS is quite dominant here, and those machines have amd hardware, and more important, next gen are going to. Its not uncommon to advocate the cheaper and more accessible solution.
This is what i heard people saying on other forums. Its nothing uncommon, everyone has their own preferences and platforms, if not then we wouldnt have any discussions like this. This whole RTX thing aint perfect either, far from it, but what hardware is.
So nothing wrong with the forum or its users, just a user base unbalance and thats what people notice.
Last edited:
DavidGraham
Veteran
Is MineCraft graphisc the same as Battlefield 1? the polygonal complexity difference alone is staggering. Not to mention physics, dynamic lights, dynamic shadows, particles .. etc.
Now why doesnt that suprise me
Must be a conspiricy that were just not seeing bfv rt on it, or Quake 2?Does nvidia pay devs to withold them from enabling RT in games? Sony aaa games lacking RT is strange too, they shouldnt be limited by contract deals.
Yes and No. There were industries that did it because statistical probability/forecasting was their main focus. I'd probably say that the explosion of ML because storage is cheap, abundant and now processing power is leading to its usage everywhere especially in smaller niches where this type of AI would be unworthy of the ROI of investment back in the day.
But I mean that seems off-topic to where your original point is. The counter argument that you postulated only works in hindsight. In hindsight what if we started with compute shaders, is just about every single reason under the sun that a lot of projects don't get started. There's always this notion that someone thinks something better is going to come along. But in most cases, iterative progression is what leads us to where we want to go. And iteration is what got us to compute shaders. Iteration is progress. And like all things, iterations have to start somewhere even if it's not where you want to begin. We only know what we know now. It's not like someone was dreaming up how to build Compute pipeline both architecturally and the shader languages and the APIs all at once back when Voodoo cards were released.
Well...ok if you're going to put it like that. Doesn't feel good to read that. Its not the type of response i was expecting.
But I mean, upon this information; You should ask your friends in your nvidia circle why they even went this route considering their investment in VXGI and VXGI2.0 - and all their other work with Gameworks that is basically the compute versions of these features.
The question is: Which game has more realistic lighting? Minecraft has shown the most accurate lighting in this threat. It's the only one shown which has infinite bounces. (Q2 has just one.)
It this regard Minecraft is ground truth. Ignoring the spatial limitation of surface cache (which is not visible to me), a offline pathtracer rendering for hours could NOT do better! It is the most impressive lighting shown here.
What remains incorrect is likely a perfect diffuse material which does not exist in reality, but if there would be perfect cubes in reality, they would look exactly like this.
But if you turn the lack of geometric detail against it, then i could do just the same with Quake 2 RTX. It will drop performance a lot if you try this within BFV, i'm sure of. Or do you believe lies like 'geometric complexity does not matter'? And you need at least 5 bounces to be realistic, not one.
I can only do the Q2 stuff, not BFV! Only triangle raytracing can show exact reflections of triangles, but i use discs to approximate stuff.Now why doesnt that suprise me
The reason you did not see this on PS4 would be: I'm coding too slowly. (In this sense i better close the page for today now... ;D )
There is a fundamental difference between these examples and games.
The examples have a very specific algorithymic computation goal that is set in stone, and devs are looking for the fastest path to run that one algorythym as fast as they can.
Games are not like that. Game's goal is the best approximation of a certain look, but the algo to get there is always up for changing. Nothing is set in stone. It's a tight rope of ambitions and compromises. So fixed function HW is always a bit short sighted, because it focus on today's algorithmic solutions, which are not always (actually never are) tomorrow's problems. Now, I'm not against all forms of HW acceleration for current trends, I just think it's necessary to exercise caution on how specific and limiting such architectural choices are.
That's a chicken and egg issue really. Your'e going to select the hardware option that is going to hit the most cases and easiest to deploy even if it's not the most efficient or the best. But accelerating it is going to outperform the ones that aren't accelerated. Every company is going to have to take a leap of faith and hope theirs is going to win. If no one makes the leap of faith, we won't have this conversation at all.
It's easy to look back in hindsight about what could have been. But back then, games industry wasn't nearly as big as it is today. The architectures not as complex. The amount of silicon available and the amount of transistors were limited compared to what came out at the time.
That what we needed at the time. The idea that we could have invented compute programming back then is so unfathomable, it doesn't even make sense; we can't look backwards saying we could have started where we are today without looking at all the limitations we had back then that drove those decisions in the first place. You can only do what you can that is best for the near future.
... compiling, time for another post...
Yes, i guess 7970 is almost twice PS4 GPU. I have 5870 (or 50) which is a bit larger as well, and i used it as reference. But now i target next gen, i'm too late. (And of course that's just a claim from a random internet stranger.)
Exactly, and we've been there, and graphics programming has matured substantially, and it's taken 20 years to do so. And besides sheer bruteforce improvements granted by microchip tech evolution, one of the biggest reason game graphics improved so much were clever Dev aproximations enabled by hw that's more programmable at every gen. The March of progress has been hand in hand with flexibility. Going to overly specific FF solutions goes against the current that has given us the most progress recently.
Next gen as in PS4?
Next gen means most likely Navi in PS5 and next XBox. It also beans i have to add sharp reflections - my blurry stuff is no longer good enough.
I assume AMD will present a very good chip for this reason, as they did with GCN. Unlike NV they do 'revolutions' at very low frequency, but they make very good and long standing chip design. They keep alive by destroying the competition only once or two times in a decade, and then they stick at the architecture for longer.
I'm pretty sure Navi will have raytracing, but no clue how it will look like. I do not expect it will beat Turing RT performance, but at least on consoles they can expose everything. And if it's more flexible, it may be the right thing with better end results.
Similar threads
