Next Generation Hardware Speculation with a Technical Spin [2018]

Status
Not open for further replies.
Could you please expand on that? I'm not very knowledgeable on the matter.

Is it because Sony is partnering with AMD and developers beat mostly to the drum of consoles?

Is MS sure to partner up with AMD as well?

Thanks.


This subforum is about consoles, so the only AMD vs nvidia optimization that could happen would be between the Nintendo Switch and all the others, who are on a completely different league in power and developer investment.
I. e. it doesn't apply.

For the PC, what you suggested is exactly what has been happening for the past 15-20 years, but the discussion isn't relevant for the thread.
 
That's not in doubt. The question is how much better is than just compute? Looking at Turing, AFAICS the Tensor and BVH blocks are about 50% of the chip, which tallies with the die size versus core count. How much collision detection, audio simluation, and whatnot could be performed on that as just compute? Would it be half as much? 20% less? The same amount?
This is the dieshot NVIDIA has provided for TU102:
upload_2018-10-22_0-35-12.png

Each of those visible SM units has bunch of Tensor-cores and 1 RT core in them. I seriously doubt they'd take even half a SM unit, which would still leave it under 50% of whole chip
I doubt they're even 33% of the chip
 
Much better. Chaosgroup gets 60% more performance from their first implementation: https://www.chaosgroup.com/blog/profiling-the-nvidia-rtx-cards
60% more performance using 50% of the silicon isn't much better, but marginally better. If everything scales accordingly, replace the RT stuff in 2080 with CUDA cores and you'll double up on them, meaning the CUDA score would be half the time as that recorded in these experiments. In fact, some of the examples show CUDA cores are faster per mm² of silicon than RT. AI33_008 renders in 652 seconds on CUDA, 465 on RTX. If 2080 GTX had twice as many CUDA cores, it should render in significantly under the RTX time.

The best pro-RTX metric out there is the gigarays-per-second metric, with the 2080 GTX being 8x the 1080, so perhaps 4x faster using the above back-of-an-envelope considerations. As I've mentioned before though, I highly doubt this is an optimal solution. But it's still a consideration, that by going with nVIdia style RT hardware, future software may be held back. What if devs develop a better cone/traced solution that doesn't need the denoising that runs on compute? Half you silicon tracing noisy rays would then look as backwards as having bitmapped graphics did when trying to render 3D spacecraft. At least with just going with generic compute, you have all doors open, even if some are slower than the other option.
 
Software-wise, what they've been doing with Radeon-Rays is quite interesting, though.

https://pro.radeon.com/en/software/radeon-rays/
I would like to see some RTX cards running Radeon-Rays, since it's not exclusive to AMD hardware. :D

I wonder if we will know soon if AMD will eventually include some kind of HW support on their cards in the near future.

FWIW, Radeon Rays leverages Intel's embree, both of which Simon Pilgrim from Sony follows on github.
 
Good post, except missing a significant perspective that contrasts this theory.

If you throw a neural net at antialising a black line on a white background, you could train and execute a model that could get great results, but if you use Xiolin Wu's algorithm, you can draw that line perfectly with very little processing requirement, because the problem can be distilled down to a mathematical solution.
I think this is one of those scenarios where you don't use a nuclear power plant to boil a cup of tea. Using the right tool for the job still applies.
It's not at all given that ML is the best way to solve upscaling. No-one's performed a decent comparison of nVidia's system versus Insomniac's, and Insomniac's upscaling is running on a poky APU after rendering the entire game. There's a good case to be made that conventional algorithm based upscaling on compute could get the same results (or better) as ML based solutions with a fraction of the silicon cost.
It's not, but it's certainly very effective and I'm willing to bet with Tensor cores it would be extremely effective.
Which one of the RTX demos wasn't using nVidia's denoising but was running it on compute? The fact I can't readily remember shows it didn't look substantially worse than the ML solution.
All the RTX demos were using algorithmic denoising. Nvidia only showcased AI denoising at Siggraph for Optic. No DirectML, no realtime neural networks.

So perhaps given two GPUs with 300 mm², on one you'd have 150 mm² of compute and 150 mm² of machine learning and BVH for noisy tracing followed by ML denoising and upscaling, and on the other you'd have 300 mm² of compute that is slower at the ray tracing but solves lighting differently and upscales with reconstructive algorithms, and the end results on screen may be difficult for people to judge.
Perhaps. But you haven't seen realtime ML denoising coupled with ML AI-Up-Resolution yet. No one has, because that's a tremendous amount of effort on both drivers and an API that isn't ready for prime time yet. Discounting it before it's even ready is fair, because it's not out, but 2 years from now I'm not certain this will hold.

A GPU 2x the power of PS4 (3.6 TF) could render a PS4-quality game (HZD, Spider Man, GoW) from a secondary position to create reflections, which would look a lot closer to RT'd reflections and not need RTing or denoising at all... ;) And before the pro-RT side says, "but that perspective wouldn't be right," of course it wouldn't, but the question is would it be close enough? Side by side, RT console versus non-RT console, would gamers know the difference? Enough to buy the RT console, all other things being equal?
We have games with planar, cubed, SSAO etc type reflections and we've seen a game with voxel cone tracing, I'm not sure what you're trying to imply here. Do they work?
Sure.
But we're talking about getting to next generation graphics not just more of the same right?
When you compare the reflections in this demo with RTX on vs when RTX is off. It's such a massive difference right here
And when it's off, that's what we're used to seeing. We're just so used to accepting low quality reflections and shadow maps your brain is probably outright ignoring it.
We're so used to baked lighting and non GI solutions, that we probably think it looks correct.

Once you get used to correctness, I don't think you'll be able to go back.

This is a very similar debate to when a user switches to 4K gaming. I had everyone under the moon tell me that 1080p and 4K you couldn't tell the difference. But once you get used to 4K or even F4K, you can't go back to 1080p. It's murder.

And it's exactly the same with HDR, you don't have an HDR screen, you don't know what you're missing.
Same thing with frame rates, you don't have the frame rate, you don't know what you're missing.

So I get that it's easy to take that position that you can't notice these differences.

Except.

If you had 4K HDR Ray Traced at great FPS.
All these little things that people can't seemingly notice, would become such an apparent leap in graphical fidelity.
With 4K and HDR, you're talking about insane amount of detail since the individual pixels can pop on their own. Particle effects and everything just have so much great detail. Now we can get more AO, reflections and shadows at significantly higher quality, working together with the 4K and HDR? My mind is already blown just thinking about what that might look like. I can't change your mind on what constitutes as looking better. Just like i can't change some people's minds that 4K looks better than 1080p. I don't have a counter argument because you've already made up your mind on the topic - and you've anchored yourself on some really early pre-release footage using incomplete drivers and APIs on a less than ideal viewing setup.

We're still very early on in RT. And as you have known with consoles, launch titles are not very representative of the graphics near the end of the generation. No reason to treat this any differently.
 
Last edited by a moderator:
You've repeatedly said that the first demos of RT aren't the best it can do and it will improve as devs get used to it. That certainty is built on the same concepts by which we know new algorithms will appear, because devs always find better ways to use hardware and always find improving algorithms.

In the long run, we won't have ray tracing. Ray tracing is just a means to sample volume. A more generalised approach to solving that problem, like getting rid of Vertex and hull shaders, and domain and geometry shaders, and replacing them with task and mesh shaders, which are more generalised.

That's not in doubt. The question is how much better is than just compute? Looking at Turing, AFAICS the Tensor and BVH blocks are about 50% of the chip, which tallies with the die size versus core count. How much collision detection, audio simluation, and whatnot could be performed on that as just compute? Would it be half as much? 20% less? The same amount?

That's the real question.
1) There's a difference between expecting an existing technology to improve and expecting a revolutionary technology to appear out of nowhere.

2) Do you mean we won't have RT-specific hardware? If you mean ray tracing per se then I disagree. Every high-end CG renderer is based on path tracing nowadays.

3) We'll have to wait for numbers for that.

Yes, but the point I'm making is that more flexible but slower RT is likely preferable to inflexible but faster RT. Right now, RTX fits into the latter category. Purely compute based solutions lie at the far end of the first category.

What would be good for consoles isn't what RTX is offering it's something more towards the flexible end of the spectrum if it is slower at RT, as long as it isn't also slower in general compute and rendering.

And when I say slower/faster than RTX, I'm talking in terms of performance per mm^2. No console maker with any bit of sanity is putting in a monolithic die even remotely as large as a RTX 2070, for example.

Right now RTX is massively slower in terms of perf/mm^2 than the previous generation GTX cards of an equivalent tier. That wouldn't cut it for consoles.

The future may be RT, but the next console generation isn't it. The next console generation will likely still be mostly traditional rendering with some forays into RT. Going with fixed function RT that can't be used for anything else, would be console suicide.

What's needed for next generation is an architecture that can use its resources for either RT or general rendering or both. Especially as RT algorithms are likely to change and evolve over the course of the consoles lifetime.

A console that is meant to last 5+ years would be committing RT suicide by adopting fixed function RT hardware, IMO.

That's fine for PC where you can easily swap out video cards as algorithms change, but you can't do that with a console.

Regards,
SB
I disagree completely. If you look at modern game renderers they all pretty much have the same features, same thing will happen next-gen. That means whatever RT techniques are developed will be used across the board. Exotic renderers will be rare. Only very few would benefit from a slower but more flexible architecture this early on. And as I said before, you can always supplement with compute for special cases.
 
I disagree completely. If you look at modern game renderers they all pretty much have the same features, same thing will happen next-gen. That means whatever RT techniques are developed will be used across the board. Exotic renderers will be rare. Only very few would benefit from a slower but more flexible architecture this early on. And as I said before, you can always supplement with compute for special cases.

I'm not even talking about exotic renderers, but the fact that the top AAA engines often go through massive changes between the start and end of a console generation. They can do that due to the flexible nature of modern GPUs and in the PS3's case, the cell SPUs made up for the massive shortcomings of the GPU.

The next console generation, if one console goes with fixed function hardware that is faster at RT at the start of the generation versus a more general purpose GPU that is slower at RT at the start of the generation, it wouldn't surprise me at all if the second one ended up faster at RT by the end of the generation than the first one.

5-8 years is a long time to be stuck with stagnant (fixed function) RT hardware if it ends up being non-optimal for the most efficient RT algorithms.

Right now, no one knows what will end up being the most optimal way to do RT in games. Betting on fixed function hardware is like flipping a coin. If you win, good for you. If you lose, you just consigned yourself to the bottom of the barrel, because there isn't anything you can do about it.

Regards,
SB
 
This subforum is about consoles, so the only AMD vs nvidia optimization that could happen would be between the Nintendo Switch and all the others, who are on a completely different league in power and developer investment.
I. e. it doesn't apply.

For the PC, what you suggested is exactly what has been happening for the past 15-20 years, but the discussion isn't relevant for the thread.

I'm confused.

I am still talking about consoles and to be more specific, is it a certainty that MS will surely not partner up with Nvidia at least with the GPU side of things?

It'll be relevant to mass adoption of Nvidia's hardware designs if it's able to get into at least the next gen Xbox.
 
I'm confused.

I am still talking about consoles and to be more specific, is it a certainty that MS will surely not partner up with Nvidia at least with the GPU side of things?

It'll be relevant to mass adoption of Nvidia's hardware designs if it's able to get into at least the next gen Xbox.

Highly unlikely. NV aren't terribly interested in low cost, low margin designs. The Switch was an interesting case where they dumped excess stock of a product not many companies wanted.

Conversely, MS (or Sony for that matter) aren't interested in paying for a high cost SOC. The console space is far too price sensitive for that, and neither wants to sell their console at a large loss at the start of the generation much less though the entire generation.

Regards,
SB
 
Highly unlikely. NV aren't terribly interested in low cost, low margin designs. The Switch was an interesting case where they dumped excess stock of a product not many companies wanted.

Conversely, MS (or Sony for that matter) aren't interested in paying for a high cost SOC. The console space is far too price sensitive for that, and neither wants to sell their console at a large loss at the start of the generation much less though the entire generation.

Regards,
SB

Thanks.

This sounds good for AMD at least.

I wonder how much loss they'd be willing to take per hardware in order to gain mindshare fast specially early on if hardware featureset/power is really a huge impact on purchasing decision.

This gen had the whole MS no used physical games fiasco, the initial powergulf (30-40% more resolution?) plus $ 100 difference/Kinect all against the earliest moves of MS.

A lot of "technical reviews" were pointing towards Sony thanks to the resolution wars.

If MS goes with Nvidia, I wonder how much marketable it would be brag about the "best/better GPU product line" and tout how much better their architecture, tensor cores. rt cores are compared to Sony's and AMD's offering but eat up a larger loss.

I do know that's just one factor and an also large factor would be exclusives which MS is gearing up for recently.

Then again, they could probably just ask for a better part and later release date to keep things smoother by sticking with AMD.
 
I wonder how much loss they'd be willing to take per hardware in order to gain mindshare fast specially early on if hardware featureset/power is really a huge impact on purchasing decision.
nVidia doesn't need mindshare...
 
if one console goes with fixed function hardware that is faster at RT at the start of the generation versus a more general purpose GPU that is slower at RT at the start of the generation, it wouldn't surprise me at all if the second one ended up faster at RT by the end of the generation than the first one.
The reason why RT has fixed function acceleration in the first place is that generic hardware is just too slow to do the job. Many have tried and the results are just not good enough. There are some parts of the graphics pipeline that are just not suitable for generic hardware. You can't do without ROPs or Texture units (Intel tried and failed), you can't do without a hardware Tessellator as the software one won't be fast enough. You can't do without a Rasterizer. I argue that you can't do the same with RT without some sort of hardware acceleration, proprietary or otherwise.
 
If MS goes with Nvidia, I wonder how much marketable it would be brag about the "best/better GPU product line" and tout how much better their architecture, tensor cores. rt cores are compared to Sony's and AMD's offering but eat up a larger loss.
The only justification for Microsoft going with NVIDIA the next generation would be to leapfrog Sony on the GPU front. As NVIDIA has better performance per watt and better performance per mm, as well as the faster architectures on PC. But that doesn't necessarily translate to a better custom GPU design for consoles. It might not matter at all.
 
Last edited:
amd based
30TF uarch after navi
1TB ssd
4TB hdd
64gb gddr6
$399
1Q 2019

NVidia
1080 based
8c cpu
16GB gdd5x
1TB hdd
$399
2020

And I'd still put money on it being AMD, that's how much chance I'd give of either Sony or MS going NVidia.
There's so many reasons why people don't give a second thought to it being NVidia, that it's pretty much just excepted.
(this time next week we hear it will be NVidia based :runaway:)
 
The reason why RT has fixed function acceleration in the first place is that generic hardware is just too slow to do the job. Many have tried and the results are just not good enough. There are some parts of the graphics pipeline that are just not suitable for generic hardware. You can't do without ROPs or Texture units (Intel tried and failed), you can't do without a hardware Tessellator as the software one won't be fast enough. You can't do without a Rasterizer. I argue that you can't do the same with RT without some sort of hardware acceleration, proprietary or otherwise.

Not sure of this, Mesh and tak shader will do a better job because it is more flexible. And a compute adaptative tesselation shader will do a better job too....



Adaptative tesselation with compute shading or mesh shading:

 
Not sure of this, Mesh and tak shader will do a better job because it is more flexible.
That maybe true, but how was this even possible in the first place without a hardware design that accommodates these shaders? You can think of Task and Mesh shaders as a hardware accelerated compute solution for Geometry. Without them, running Tessellation on regular compute units would a performance suicide.
 
Status
Not open for further replies.
Back
Top