Titan V ray tracing (almost) like a RTX 2080 TI

[ne0tic]

I’m really looking foward to seeing how the Ray Tracing performance will be when games actually take advantage of the RT cores! 

[Mira Yurizaki]

I posted this to another thread but I think the missing information that people aren't aware of is DXR has a fallback compiler mode. https://github.com/Microsoft/DirectX-Graphics-Samples/tree/master/Libraries/D3D12RaytracingFallback. Some people on Reddit compiled some samples Microsoft provided. What's really interesting is that someone posted results from a Titan V and someone else posted results from an RTX 2080. They both performed more or less the same as far as rays per second goes. And this was months ago.

 

So if anything, the theory would be that Battlefield V is really built using the fallback layer. One could argue that Dice, because possibly through an exclusive deal with NVIDIA, arbitrarily restricted DXR features to RTX cards. However I would argue that DXR isn't a free lunch that you can enable just because you're using a DX12 card and Windows 10 1809, it still requires driver support. Much like how first generation DX11 cards are technically compatible with DX12 because DX12 supports FL11_0, which is the feature level introduced in DX11. But because there's no DX12 driver support, those cards can't run DX12 games.

 

EDIT: Now that I think about it, I think you can run those demos on Maxwell and Pascal, so yeah, Dice probably arbitrarily restricted DXR to RTX cards.

[S w a t s o n]

16 minutes ago, Mira Yurizaki said:

I posted this to another thread but I think the missing information that people aren't aware of is DXR has a fallback compiler mode. https://github.com/Microsoft/DirectX-Graphics-Samples/tree/master/Libraries/D3D12RaytracingFallback. Some people on Reddit compiled some samples Microsoft provided. What's really interesting is that someone posted results from a Titan V and someone else posted results from an RTX 2080. They both performed more or less the same as far as rays per second goes. And this was months ago.

 

So if anything, the theory would be that Battlefield V is really built using the fallback layer. One could argue that Dice, because possibly through an exclusive deal with NVIDIA, arbitrarily restricted DXR features to RTX cards. However I would argue that DXR isn't a free lunch that you can enable just because you're using a DX12 card and Windows 10 1809, it still requires driver support. Much like how first generation DX11 cards are technically compatible with DX12 because DX12 supports FL11_0, which is the feature level introduced in DX11. But because there's no DX12 driver support, those cards can't run DX12 games.

 

EDIT: Now that I think about it, I think you can run those demos on Maxwell and Pascal, so yeah, Dice probably arbitrarily restricted DXR to RTX cards.

Have to disagree, if they are using the fallback layer it would be purely shader driven and 2080 aint gonna keep up with Titan V

[Mira Yurizaki]

6 minutes ago, S w a t s o n said:

Have to disagree, if they are using the fallback layer it would be purely shader driven and 2080 aint gonna keep up with Titan V

That was based on the fallback compiler application, as two Reddit users posted these:

Quote

BigJon901

 

2 points · 3 months ago

 

Titan V, not overclocked, watercooled.

First test, the reflections one:

• 376 Million Primary Rays/s, 46fps, 21ms (4k fullscreen)

• 260Million Primary Rays/s, 255fps, 3.60ms (default screen size)

Second test, the cube one:

• 8484 Million Primary Rays/s, 1058fps (4k fullscreen)

• 3902 Million Primary Rays/s, 4200fps (default screen size)

 

Quote

 

level 1

Zelminar

 

3 points · 3 months ago · edited 3 months ago

 

GTX 2080

i7-7700k @4.20Ghz

Reflection/Geometry Test:

• 3840x2065 (~4k windowed w/title) 2025mhz - https://i.imgur.com/NoQJAKD.jpg 397 Million/s, 114 fps, 8.87ms

• 3840x2065 (~4k windowed w/title) 1800mhz - https://i.imgur.com/cbqL7wz.jpg 369 Million/s, 103 fps, 9.53ms

• 1080p (windowed) 2025mhz - https://i.imgur.com/8B8YwCj.png 310 Million/s, 320 fps, 2.97ms

• 1080p (windowed) 1800mhz - https://i.imgur.com/4g5KtFA.png 289 Million/s, 275 fps, 3.19ms

Cube Test:

• 3840x2065 (~4k windowed w/title) 2025mhz - https://i.imgur.com/tF32EMA.png 6737 Million/s, 1911 fps

• 3840x2065 (~4k windowed w/title) 1800mhz - https://i.imgur.com/TPTk0Ek.png 6148 Million/s, 1744 fps

• 1080p (windowed) 2025mhz - https://i.imgur.com/gT2ViZE.png 4608 Million/s, 5000 fps

• 1080p (windowed) 1800mhz - https://i.imgur.com/M8zCfd5.png 4204 Million/s, 4561 fps

 

Also the app measures primary ray performance, but not secondaries which may affect performance. However, at the very least the point is that DXR has a fallback layer in the event that the graphics hardware doesn't have hardware RT acceleration. And while I'd like to have other tech sites take a look at this, all I really have to go on is that Reddit thread. So the only explanation I have that the Titan V can perform as well as an RTX 2080 Ti and an apparent lack of RT core usage is that DICE chose to use the fallback layer at launch for compatibility with whatever they were using for development. And I'm willing to put down money for an AAA title that's pushing the hardware envelope, getting something to work is a much higher of a priority than getting said thing to work well.

[S w a t s o n]

34 minutes ago, Mira Yurizaki said:

That was based on the fallback compiler application, as two Reddit users posted these:

 

 

Also the app measures primary ray performance, but not secondaries which may affect performance. However, at the very least the point is that DXR has a fallback layer in the event that the graphics hardware doesn't have hardware RT acceleration. And while I'd like to have other tech sites take a look at this, all I really have to go on is that Reddit thread. So the only explanation I have that the Titan V can perform as well as an RTX 2080 Ti and an apparent lack of RT core usage is that DICE chose to use the fallback layer at launch for compatibility with whatever they were using for development. And I'm willing to put down money for an AAA title that's pushing the hardware envelope, getting something to work is a much higher of a priority than getting said thing to work well.

From what I see on that reddit post those demos run in hardware accelerated mode on compatible hardware and fallback mode for non-compatible.

 

So in this case that means RTX 2080 was using RT Cores, possibly tensor. Titan V was using sheer cuda count, cpu resources and possibly tensor to come close. It's got like double the cuda cores.

 

What this means is that if BFV's raytracing is only fallback mode means the 2080 is now using like half the shaders of Titan V to ray trace and denoise. It's not gonna work very well

[S w a t s o n]

5 minutes ago, Mira Yurizaki said:

 So the only explanation I have that the Titan V can perform as well as an RTX 2080 Ti and an apparent lack of RT core usage is that DICE chose to use the fallback layer at launch for compatibility with whatever they were using for development.

My explanation is simple. Tensor cores

[Mira Yurizaki]

36 minutes ago, S w a t s o n said:

From what I see on that reddit post those demos run in hardware accelerated mode on compatible hardware and fallback mode for non-compatible.

 

So in this case that means RTX 2080 was using RT Cores, possibly tensor. Titan V was using sheer cuda count and cpu resources to come close. It's got like double the cuda cores.

 

What this means is that if BFV's raytracing is only fallback mode means the 2080 is now using like half the shaders of Titan V to ray trace and denoise. It's not gonna work very well

I'll bite that the RTX 2080 in the Reddit post was using the hardware accelerated version, but if we're going to talk about the case here with a Titan V performing the same as the 2080 Ti, then it's still plausible BFV is using the fallback layer exclusively because the raw performance figures between the two GPUs are close.

 

26 minutes ago, S w a t s o n said:

My explanation is simple. Tensor cores

That doesn't make sense when tensor cores are matrix math units. I don't see how resolving a BVH tree is done with matrix math. Besides that, the only thing NVIDIA said the tensor cores on Volta were used for was for denoising: https://www.anandtech.com/show/12546/nvidia-unveils-rtx-technology-real-time-ray-tracing-acceleration-for-volta-gpus-and-later

 

Quote

Meanwhile NVIDIA also mentioned that they have the ability to leverage Volta's tensor cores in an indirect manner, accelerating ray tracing by doing AI denoising, where a neural network could be trained to reconstruct an image using fewer rays, a technology the company has shown off in the past at GTC

 

[mr moose]

On 1/4/2019 at 3:30 AM, S w a t s o n said:

 

2080ti and Titan V share similar number of tensor cores hence similar AI performance.

 

Titan V has 640 while the 2080 ti has 544, which is a fair difference, but lacking any technical knowledge in how they render RT on the two cards maybe the extra tensors make up for the lack of RT cores seeing as the tensor cores seem to be at the base of RT?

[S w a t s o n]

1 hour ago, Mira Yurizaki said:

I'll bite that the RTX 2080 in the Reddit post was using the hardware accelerated version, but if we're going to talk about the case here with a Titan V performing the same as the 2080 Ti, then it's still plausible BFV is using the fallback layer exclusively because the raw performance figures between the two GPUs are close.

 

That doesn't make sense when tensor cores are matrix math units. I don't see how resolving a BVH tree is done with matrix math. Besides that, the only thing NVIDIA said the tensor cores on Volta were used for was for denoising: https://www.anandtech.com/show/12546/nvidia-unveils-rtx-technology-real-time-ray-tracing-acceleration-for-volta-gpus-and-later

 

 

Ok so the actual way it works is that it takes an image with a low number of rays traced (something like single pass iirc?) and infers what the image would look like after thousands of iterations. While tensor cores are "denoising" it is in fact infering the additional rays. You are right that's not actually resolving the BVH, that's why RT Cores exist. It's not denoising in the way DICE is. "Denoising" in raytracing is effectively just tracing additional rays. DICE are tracing a given number of rays (single pass + denoise) and then passing it by a temporal filter which is just taking the actual calculated (including inferred) rays and basically uses advanced custom AA to clean it up. They call it denoise but it's not the same thing as what the tensor cores are doing. One is akin to actually  tracing the rays and one is akin to antialiasing. So you are right to call it denoising but my point about it being tensor cores is still valid.

 

So current theory:

 

BFV RTX Behaviour: The cards are taking a real single pass of raycasting (RIP shaders), inferring a high number of rays per pixel (actual denoising) and then they get the temporal filter. (fake "denoising" super fancy AA just for the raytracing)

 

Expected full RTX implementation: RT Cores produce real raytraced rays at a high ray per pixel count (shaders aint mad and now the lighting look way better) and then it's going to infer an even higher ray per pixel count using tensor cores. (actual denoise on like 10x the amount of source rays)

 

This is how tensor denoise works.

 

 

45 minutes ago, mr moose said:

Titan V has 640 while the 2080 ti has 544, which is a fair difference, but lacking any technical knowledge in how they render RT on the two cards maybe the extra tensors make up for the lack of RT cores seeing as the tensor cores seem to be at the base of RT?

I wouldn't call it the base of RT. It's just that some clever nvidia researches figured out how to use an AI model to infer rays based on a model. They published a white paper using Volta which for whatever reason I can no longer find, basically stating that it was not ready for use at the time because frame times weren't low enough. This was basically the first mention of RTX. They developed RT Cores for a reason, they are so much faster. Tensor is just used for denoise and DLSS so it's always part of RTX even if not to calculate ray intersects.

[leadeater]

6 minutes ago, S w a t s o n said:

Tensor is just used for denoise and DLSS so it's always part of RTX even if not to calculate ray intersects.

And without Tensor cores we would not be able to have playable games with good enough frame rates. RT cores will be able to, or should be able to, increase the performance of the ray casting or allow for more rays but it's still not going to be enough to produce a clean image. Both RT cores and Tensor cores are about as important as each other.

[S w a t s o n]

10 minutes ago, leadeater said:

And without Tensor cores we would not be able to have playable games with good enough frame rates. RT cores will be able to, or should be able to, increase the performance of the ray casting or allow for more rays but it's still not going to be enough to produce a clean image. Both RT cores and Tensor cores are about as important as each other.

True, it's a weird relationship though. Volta is slow at raytracing in comparison to turing because of that single pass on the shaders but the end result is not bad because of Tensor Cores. RT Cores are basically just there so the shaders dont get literally rekt. There could come a point where they only include enough RT cores for a few rays per pixel and just smash more tensor cores on to infer to crazy amounts as the source data will be good enough and fast enough. Even RT Core doing 1 ray per pixel would be ok seeing as volta makes it past quality check with a single pass. As long as the shader doesnt have to do it basically

[CarlBar]

On 1/3/2019 at 3:23 AM, S w a t s o n said:

Wrong, literally wrong.
https://nvidianews.nvidia.com/news/nvidia-rtx-technology-realizes-dream-of-real-time-cinematic-rendering

 

You whole argument is wrong, it can infer rays but it's not faster than an ASIC circuit designed to calculate them specifically. They specifically state they are not using RT cores. If you think they are not using tensor cores and they ADMIT they are not using RT cores do you think they are magic?
 

 

Apologies for slow get back, sleeping all kinds of funny.

 

That doesn't say what you think it says. At no point does it say the Tensor Cores do the raycasting, (note i'll comment on your later comments on this), it say they can do offload. Which they can. Via doing the denoising. It says nothing there about Tensor cores specifically being able to do raycasting.

 

Second, your proposed scenario, (using the tensor cores to extrapolate additional rays from a handful of samples), is perfectly doable. It's also known as Tensor core denoising. Which we know DICE is NOT doing. For yuor scenario to make sense they have to be casting the rays with the Tensor cores.

 

Also the gap in performance between an RTX 2080Ti and a Titan V in CUDA isn't as great as you seem to think. the 2080Ti has 90% the CUDA cores at about a 6% higher clock speed, there's only a few percent advantage to the Titan V.

 

Lastly your pascal comparisons are meaningless because there;s more differences than just the new hardware between the two, (the main one is the 60 times greater half precision jump, if half precision is heavily used in fallback denoising then it's naturally going to produce massive gains in performance).

[S w a t s o n]

2 hours ago, CarlBar said:

 

Apologies for slow get back, sleeping all kinds of funny.

 

That doesn't say what you think it says. At no point does it say the Tensor Cores do the raycasting, (note i'll comment on your later comments on this), it say they can do offload. Which they can. Via doing the denoising. It says nothing there about Tensor cores specifically being able to do raycasting.

 

Second, your proposed scenario, (using the tensor cores to extrapolate additional rays from a handful of samples), is perfectly doable. It's also known as Tensor core denoising. Which we know DICE is NOT doing. For yuor scenario to make sense they have to be casting the rays with the Tensor cores.

 

Also the gap in performance between an RTX 2080Ti and a Titan V in CUDA isn't as great as you seem to think. the 2080Ti has 90% the CUDA cores at about a 6% higher clock speed, there's only a few percent advantage to the Titan V.

 

Lastly your pascal comparisons are meaningless because there;s more differences than just the new hardware between the two, (the main one is the 60 times greater half precision jump, if half precision is heavily used in fallback denoising then it's naturally going to produce massive gains in performance).

Cuda core comparison was 2080 and titan v not 2080ti as that was your comparison on reddit thread. If you read my other responses you'd see I mentioned all this but they were long posts so ok.

 

In regards to the fact that DICE said they were not running tensor core denoise. I think they meant DLSS or otherwise at this point because there is no chance they are running on fallback layer. Microsoft themselves have said it was only developed to test cards that had NO hardware acceleration and will not be further developed or updated. Again if they are running on fallback layer, 2080 just wont even be close to running BVF RT like Titan V let alone poor little RTX 2060. The tensor cores are not traversing the BVH to actually raycast, but they are inferring what the correct rays that SHOULD be cast.

 

I am 99% sure the tensor cores are being used to infer higher ray count from a single pass done on the shaders.

[CarlBar]

3 hours ago, S w a t s o n said:

Cuda core comparison was 2080 and titan v not 2080ti as that was your comparison on reddit thread. If you read my other responses you'd see I mentioned all this but they were long posts so ok.

 

In regards to the fact that DICE said they were not running tensor core denoise. I think they meant DLSS or otherwise at this point because there is no chance they are running on fallback layer. Microsoft themselves have said it was only developed to test cards that had NO hardware acceleration and will not be further developed or updated. Again if they are running on fallback layer, 2080 just wont even be close to running BVF RT like Titan V let alone poor little RTX 2060. The tensor cores are not traversing the BVH to actually raycast, but they are inferring what the correct rays that SHOULD be cast.

 

I am 99% sure the tensor cores are being used to infer higher ray count from a single pass done on the shaders.

 

DLSS is completely seperate (in terms of conceptuality and graphics settings), from denoising, they even showed an image of what they where running their custom denoising on and it's clearly a pre-denoised raytrace.

 

The more i hear about BFV RTX the more i think it was DICE and NVIDIA pulling a marketing stunt. I don't think we''ll see a true RTX implementation for a few more months when devs who weren't quite as deep into development are able to do a full implementation.

 

Also i didn't mention any reddit thread. That was another poster. And BFV on the Titan back at the showoff and first release BFV RTX on a 2080Ti ran virtually the same frame rates with the same apparent quality.

 

Like i said though turing has massively increased shader pipeline half precision. That means it can do everything the tensor cores do, but at only a couple of tens of percent the speed., (the tensor performance seems to be 3-6 times faster depending on workload),  so depending on how the denoising algorithm works it could reasonably be run. Remember the BFV denoising breaks down in some places, that suggests a much less capable denoiser than what NVIDIA claimed the Tensor cores could do, which fits with the idea that they're running it on less processing power and having to cut corners to make it work.

 

EDIT: As an aside i'm going to be very interesting to see what numbers we get out of the 3DMark Port Royal Benchmark tomorrow.  it's still going to have some flaws, (Futermarks stance on some things means NVIDIA won't be able to improve the denoising via driver releases of machine learning created enoising setups, which again raises the question of if it will even use the Tensor Cores). But it will be the second RTX enabled program and a lot more repeatable.

[S w a t s o n]

14 minutes ago, CarlBar said:

 

DLSS is completely seperate (in terms of conceptuality and graphics settings), from denoising, they even showed an image of what they where running their custom denoising on and it's clearly a pre-denoised raytrace.

 

The more i hear about BFV RTX the more i think it was DICE and NVIDIA pulling a marketing stunt. I don't think we''ll see a true RTX implementation for a few more months when devs who weren't quite as deep into development are able to do a full implementation.

 

Also i didn't mention any reddit thread. That was another poster. And BFV on the Titan back at the showoff and first release BFV RTX on a 2080Ti ran virtually the same frame rates with the same apparent quality.

 

Like i said though turing has massively increased shader pipeline half precision. That means it can do everything the tensor cores do, but at only a couple of tens of percent the speed., (the tensor performance seems to be 3-6 times faster depending on workload),  so depending on how the denoising algorithm works it could reasonably be run. Remember the BFV denoising breaks down in some places, that suggests a much less capable denoiser than what NVIDIA claimed the Tensor cores could do, which fits with the idea that they're running it on less processing power and having to cut corners to make it work.

 

EDIT: As an aside i'm going to be very interesting to see what numbers we get out of the 3DMark Port Royal Benchmark tomorrow.  it's still going to have some flaws, (Futermarks stance on some things means NVIDIA won't be able to improve the denoising via driver releases of machine learning created enoising setups, which again raises the question of if it will even use the Tensor Cores). But it will be the second RTX enabled program and a lot more repeatable.

I am aware of what DLSS is but DICE is the one calling their temporal filter denoising. Which is not denoising. I am suggesting DICE just used the wrong terms in that description.

Denoising is hard term to use correctly because in reality denoising is just raytracing. When you denoise a raytraced image you just trace more rays (additional rays per pixel or "passes"). Tensor core denoise is an inference of these rays rather than calculating them. DLSS is also just an inference model.

 

DICE is using temporal AA basically which is why the denoising is not as good, this is a known factor.

Turing SM's are not magic they are still just fp32 units but yes they are able to now process things like int32 and half precision more efficiently. Half precision capacity is still rated at only double single precision.
https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/
 

[CarlBar]

23 minutes ago, S w a t s o n said:

I am aware of what DLSS is but DICE is the one calling their temporal filter denoising. Which is not denoising. I am suggesting DICE just used the wrong terms in that description.

Denoising is hard term to use correctly because in reality denoising is just raytracing. When you denoise a raytraced image you just trace more rays (additional rays per pixel or "passes"). Tensor core denoise is an inference of these rays rather than calculating them. DLSS is also just an inference model.

 

DICE is using temporal AA basically which is why the denoising is not as good, this is a known factor.

Turing SM's are not magic they are still just fp32 units but yes they are able to now process things like int32 and half precision more efficiently. Half precision capacity is still rated at only double single precision.
https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/
 

 

Um i think your so busy overthinking what denoising is your missing somthing important.

 

First and foremost DICE said they are using a variant of their AA algorithm, not the full unmodified one. Important distinction.

 

Second, denoising is not in any way shape or form raycasting. It's not quite AA either but it shares far more in common with AA than raycasting.

 

I strongly, strongly, strongly, strongly recommend you watch adored.tv's video on the subject TBH as he has a small mountain of info on it thats very concise and easy to understand as i get the distinct impression your seriously confused about some things.

 

What both AA and Denoising do is take a set "image" (For AA it;s a 2D pixel display image, for RTX Denoising it;s a 3D raycast render), and infer non-existent datapoint values and modify existing datapoint values in the "image" based on present values in other parts of the "image", (and for Temporal AA values from future frames). There's some significant differences in there, but what they're actually doing, (inferring and modifying values based on other values), shares some common points and some parts of the algorithms are absolutely cross applicable because they're doing the same fundamental thing, just with different data input and output locations. Thats why you can run both DLSS and Denoising on Tensor Cores as well. They have differences, but they're also doing some very similar things with very similar goals, (producing a more realistic effect than a one size fits all solution).

 

I wouldn't be surprised actually to see NVIDIA add support in the future, ethier at a hardware or software level for running parts of AA calculations on Tensor Cores to help speed things up.

[S w a t s o n]

28 minutes ago, CarlBar said:

 

Um i think your so busy overthinking what denoising is your missing somthing important.

 

First and foremost DICE said they are using a variant of their AA algorithm, not the full unmodified one. Important distinction.

 

Second, denoising is not in any way shape or form raycasting. It's not quite AA either but it shares far more in common with AA than raycasting.

 

I strongly, strongly, strongly, strongly recommend you watch adored.tv's video on the subject TBH as he has a small mountain of info on it thats very concise and easy to understand as i get the distinct impression your seriously confused about some things.

 

What both AA and Denoising do is take a set "image" (For AA it;s a 2D pixel display image, for RTX Denoising it;s a 3D raycast render), and infer non-existent datapoint values and modify existing datapoint values in the "image" based on present values in other parts of the "image", (and for Temporal AA values from future frames). There's some significant differences in there, but what they're actually doing, (inferring and modifying values based on other values), shares some common points and some parts of the algorithms are absolutely cross applicable because they're doing the same fundamental thing, just with different data input and output locations. Thats why you can run both DLSS and Denoising on Tensor Cores as well. They have differences, but they're also doing some very similar things with very similar goals, (producing a more realistic effect than a one size fits all solution).

 

I wouldn't be surprised actually to see NVIDIA add support in the future, ethier at a hardware or software level for running parts of AA calculations on Tensor Cores to help speed things up.

I'm not overthinking denoising and I have seen adoredtv's video, I watch most of them. I urge you to look at raytracing in other contexts. Denoising is raytracing. I also know that DICE is using modified AA. I already pointed all of this out in my earlier posts in thread but they are quite long so I'm not gonna quote them jus tgo look if you want to see.

 

You can run both DLSS and denoising at the same time because it's just 2 inference models instead of 1 you explained it as if it's somethign else but you just described the inference model.

[CarlBar]

My understanding is that DLSS is functioning in a 2D environment like all AA whilst denoising functions in 3 dimensions. They're very similar but also have some differences. The point is however you should be able to do denoising with a modified AA algorithm.

 

And once again denoising is not raycasting, seriously what about this is hard to understand if you have watched adoredtv's video on it, he clearly explains what casting a ray is and he clearly explains what denoising is. they're 2 completely different things that complement each other to produce a usable image at a reasonable frame rate.

 

Although to be fair i still don't get why where arguing about this. DICE have said they';re not doing denoising on the Tensor cores. I expect them to understand their own software to be able to develop it so thats good enough for me. There's nothing to argue about here for me, your wrong will continue to be wrong and will allways be wrong as far as i'm concerned.

[leadeater]

3 hours ago, CarlBar said:

My understanding is that DLSS is functioning in a 2D environment like all AA whilst denoising functions in 3 dimensions. They're very similar but also have some differences. The point is however you should be able to do denoising with a modified AA algorithm.

Even traditional denoising for Ray Tracing is just image blurring, different algorithms can be used for different blurring techniques but they are blurring none the less. That means 2D as well, you can improve the quality of denoising with things like edge detection because you can lose straight edge detail when denoising using basic blurring methods.

 

Nvidia takes other input data for their denoising among other things, which Nvidia has a few dev blogs explaining etc, but for BFV not important because it's not being used.

 

BFV denoising is more traditional method, pixel adjacency checking and blurring/blending likely with some amount of edge detection (just guessing on that part).

 

I think the lines of what denoising is, commonly known as, is getting a bit blurred now since Nvidia is introducing other methods (as well as others) which do more than just blurring/blending and look to infer more samples per pixel creating a cleaner source image, stage 2, before another denoise is passed over it, stage 3. Even with all the inferencing/guess work to increase the effective ray count and sample per pixel it's still going to be a noisy image just less so. 

[CarlBar]

16 hours ago, leadeater said:

Even traditional denoising for Ray Tracing is just image blurring, different algorithms can be used for different blurring techniques but they are blurring none the less. That means 2D as well, you can improve the quality of denoising with things like edge detection because you can lose straight edge detail when denoising using basic blurring methods.

 

Nvidia takes other input data for their denoising among other things, which Nvidia has a few dev blogs explaining etc, but for BFV not important because it's not being used.

 

BFV denoising is more traditional method, pixel adjacency checking and blurring/blending likely with some amount of edge detection (just guessing on that part).

 

I think the lines of what denoising is, commonly known as, is getting a bit blurred now since Nvidia is introducing other methods (as well as others) which do more than just blurring/blending and look to infer more samples per pixel creating a cleaner source image, stage 2, before another denoise is passed over it, stage 3. Even with all the inferencing/guess work to increase the effective ray count and sample per pixel it's still going to be a noisy image just less so. 

 

Maybe i've misunderstood a key part of scene rendering but my understanding was that the onscreen image is a 2D images taken through a viewport of somthing rendered in 3 dimensions in the scene. AA is applied to the 2D image created before it hits the screen, whilst denoising is applied to the 3D render.

 

As an aside on the Tensor core discussion, i have confirmation, (scroll down past the first few posts), that the new 3dmark port royal benchmark does not use tensor cores at all.

[leadeater]

43 minutes ago, CarlBar said:

Maybe i've misunderstood a key part of scene rendering but my understanding was that the onscreen image is a 2D images taken through a viewport of somthing rendered in 3 dimensions in the scene. AA is applied to the 2D image created before it hits the screen, whilst denoising is applied to the 3D render.

Denoising isn't much different to AA, AA more specifically looks for edges and smooths where denoising is applied over the entire image. Both work on the 2D viewport. Denoising is working on the principle that there is missing information in the render/image and AA is working on the principle of smoothing a complete image.

 

The reason it appears like denoising works on 3D is because that's actually why you use Ray Tracing in the first place and this is where the benefit is coming from, if you move the viewport the lighting and shadows is accurate to that change in perspective unlike traditional rasterization rendering methods where you get issues of pop in and pop out of reflections because the object being reflected is no longer in the viewport. Denoising is still working on a flat 2D image.

[leadeater]

32 minutes ago, CarlBar said:

As an aside on the Tensor core discussion, i have confirmation, (scroll down past the first few posts), that the new 3dmark port royal benchmark does not use tensor cores at all.

Interesting, so for 3DMark it's pretty much pure RT accelerated. Makes sense because the purpose is to benchmark ray tracing performance, not denoising, and also needs to be supported across different GPU vendors so proprietary methods are out due to that. If Nvidia were using the Tensor cores more to accelerate a DX12 denoising function, meaning any GPU vendor could, then they would probably be able to be used. 

[Stefan Payne]

On 1/1/2019 at 7:43 PM, Tech Enthusiast said:

Original posts on German forum: 3dcenter

 

pls link the posting, not the page number. Or mention the Posting number...

[CarlBar]

2 hours ago, leadeater said:

Denoising isn't much different to AA, AA more specifically looks for edges and smooths where denoising is applied over the entire image. Both work on the 2D viewport. Denoising is working on the principle that there is missing information in the render/image and AA is working on the principle of smoothing a complete image.

 

The reason it appears like denoising works on 3D is because that's actually why you use Ray Tracing in the first place and this is where the benefit is coming from, if you move the viewport the lighting and shadows is accurate to that change in perspective unlike traditional rasterization rendering methods where you get issues of pop in and pop out of reflections because the object being reflected is no longer in the viewport. Denoising is still working on a flat 2D image.

 

Ahh ok i get it now. i thought the denoise worked on where the rays hit rather than back at the viewport, (as some forms of raytracng don't go to/from the viewport but rather from a point on the render towards the light).

 

2 hours ago, leadeater said:

Interesting, so for 3DMark it's pretty much pure RT accelerated. Makes sense because the purpose is to benchmark ray tracing performance, not denoising, and also needs to be supported across different GPU vendors so proprietary methods are out due to that. If Nvidia were using the Tensor cores more to accelerate a DX12 denoising function, meaning any GPU vendor could, then they would probably be able to be used. 

 

Sure but given DICE apparently is using their own algorithm it's not clear fi DX12 actually has a denoiser of it's own, (or even if 3dmark is using it instead of their own solution if it does).

 

Also to be blunt i expect a benchmark to be representative of real world performance differences. A proper implementation of NVIDIA's RTX support involves using Tensor core denoising. If a benchmark doesn't use it it's fundamentally usless for comparing to another manufacturers solution. I'm no great fan of some of the stuff NVIDIA has been pulling, but that dosen;t mean i won't be pissy if someone biases against them either.

 

That said i can cut futuremark some slack as despite many claiming they will use the Tensor cores, we haven't seen much of anything of it being released yet, (understandably, people complaining need to look at the DX12 release schedule IMO), so we can;t confirm that will be true.

 

I also wouldn't have any objection to Futermark running their own tensor core based denoiser for control purposes.

[leadeater]

50 minutes ago, CarlBar said:

Also to be blunt i expect a benchmark to be representative of real world performance differences. A proper implementation of NVIDIA's RTX support involves using Tensor core denoising. If a benchmark doesn't use it it's fundamentally usless for comparing to another manufacturers solution. I'm no great fan of some of the stuff NVIDIA has been pulling, but that dosen;t mean i won't be pissy if someone biases against them either.

Use of it might come later when AMD has something, problem for 3DMark is that they have to make something that will run on any compliant DX12 DXR device not just one company's implementation. Otherwise you are making a tech demo not a benchmark, we have plenty of those from Nvidia already.

 

50 minutes ago, CarlBar said:

I also wouldn't have any objection to Futermark running their own tensor core based denoiser for control purposes.

That may happen, kind of depends where industry development goes. I suspect it may be one of those it's just too new issues and there is no solid understanding of how to use Tensor cores for this purpose on a algorithmic basis rather than an AI basis, which I think using that method goes against what 3DMark allows. Allowing an implementation that is backed by an external AI training cluster isn't really benchmarking the GPU if you ask me, it's still useful information but ends up being more of a software/AI benchmark. It's sort of dipping your toes in to pre-rendering, very slightly.