Why does raytracing suck so much, is it because of nvidia's bad hardware implementation?

Status
Not open for further replies.
Snowdog said:
Clearly some kind of absurd misinterpretation on your part.
Higher RT performance will require more transistors, thus more die area, and thus a higher price.
Trying to devote the same amount of die area to RT performance across a product line, would make low end cards too expensive, and high end cards under-powered.
RT performance will be segmented just like Raster performance is. It really couldn't be otherwise.
Click to expand...
Do you realize even RTX 2060 can calculate 5 gigarays/s which translates intersection of ten times the amount of rays that is needed for 4K@60Hz if all it calculated was primary rays one ray per pixel and did not have to run any shaders for each intersection? Note: this is supposed to be from real workloads. Even if we apply bullshit-correction factor into account it means calculating ray intersection is not an issue for even weakest RTX card.

Each ray intersection need to be handled by shader program and RTX effects performance is almost entirely limited by raw shader performance.
DXR by its design is fully programmable and intended to work in this way. There are no fixed programs which Nvidia could put in to hardware which would offload shaders and accelerate "RT performance" even more.
Maybe in the future most used programs will be made into templates which will be standardized and which then GPU manufacturers will be able to accelerate in hardware. For today however adding ray intersection and tensor cores is enough of a work done.

My prediction is that future Nvidia cards will have more CUDA cores per SM block to better utilize die space.
In Turing it was probably too much hassle to change SM block design in this direction so they just added RT and Tensor cores as they are to existing slightly tweaked design and called it a day. They have best gaming GPU, GPGPU and hardware accelerated DXR GPU anyway and cards need to be released to make money...
 
demondrops said:
1080ti vs 2080ti is or atleast was a 40% increase in price. prety sure u can just look at prices of both cards and see it is more or less 40% +- can only guess i 980ti -> 1080ti was more similar in price, while the 1080ti were more expensive for me it was not by alot. idk what other reason they get to near double price of same tier graphics card in 1 generation, seems rather special. at best i can get a 2080ti for 14-1500 usd as lowest price in norway... importing is pointless cause the gov. screw u over on tax and processing. that is prety much just the same as release unfortunately.
Click to expand...
RTX if anything was an excuse for price increase and not the reason

Let's see what excuse will AMD have for Navi pricing ^_^
 
XoR_ said:
Let's see what excuse will AMD have for Navi pricing ^_^
Click to expand...

They won't need one, plenty of people on these forums are now attacking anybody who wants cheaper products because "businesses exist to make money" and "prices don't live in a vacuum", etc. Bitcoin bought both Nvidia and AMD a year or more of boosted prices and revenue, and I suppose the rise of e-sports over the past 4-5 years has as well. These defensive quotes spread through social media and forums like a virus and somehow manage to normalize exorbitant pricing.

That said, with the bitcoin boom over (I think), the used GPU market is raging and prices are coming down swiftly. I'm just sitting back until I can get a reasonable upgrade for a lower price. 2080 for <$499 would suit me.
 
Mode13 said:
They won't need one, plenty of people on these forums are now attacking anybody who wants cheaper products because "businesses exist to make money" and "prices don't live in a vacuum", etc. Bitcoin bought both Nvidia and AMD a year or more of boosted prices and revenue, and I suppose the rise of e-sports over the past 4-5 years has as well. These defensive quotes spread through social media and forums like a virus and somehow manage to normalize exorbitant pricing.
Click to expand...
Navi situation is just hilarious to me because AMD fanboys were the loudest ones to criticize Nvidia for their RTX lineup pricing and so far it seems AMD will do exactly the same move but with less excuses... or more precisely with more sad excuses
 
XoR_ said:
Navi situation is just hilarious to me because AMD fanboys were the loudest ones to criticize Nvidia for their RTX lineup pricing and so far it seems AMD will do exactly the same move but with less excuses... or more precisely with more sad excuses
Click to expand...

Isn't that how capitalism works though? That's not AMD being greedy or evil, that's AMD being realistic about being a publicly traded company.

Ideally, if AMD can compete, then as nVidia loses sales they will lower prices, AMD will follow to match, and you'll get in a small price war until someone gets to the point they can't cut the price any further.

And if AMD can't compete, then nVidia stays where they are and continues to sell at high margins. And next-gen nVidia sees if they can raise those margins just a bit further while maintaining sales.
 
XoR_ said:
Do you realize even RTX 2060 can calculate 5 gigarays/s which translates intersection of ten times the amount of rays that is needed for 4K@60Hz if all it calculated was primary rays one ray per pixel and did not have to run any shaders for each intersection? Note: this is supposed to be from real workloads. Even if we apply bullshit-correction factor into account it means calculating ray intersection is not an issue for even weakest RTX card.
Click to expand...

You do realize that even the 2060 die is very Large.
You do realize that NVida has a proportional amount of RT cores at every level, Titan RTX has double the Shaders as 2070 and it has Double the RT cores as well. If all the RT cores they needed was the amount in the 2060 they would have stopped there and added more Shaders.

But of course, you think know better than NVidia does when it comes to designing RT GPUs. :rolleyes:
 
Brian_B said:
Isn't that how capitalism works though? That's not AMD being greedy or evil, that's AMD being realistic about being a publicly traded company.

Ideally, if AMD can compete, then as nVidia loses sales they will lower prices, AMD will follow to match, and you'll get in a small price war until someone gets to the point they can't cut the price any further.

And if AMD can't compete, then nVidia stays where they are and continues to sell at high margins. And next-gen nVidia sees if they can raise those margins just a bit further while maintaining sales.
Click to expand...

The problem on these forums, is that AMD and NVidia will price similarly, and NVidia gets demonized for it, and AMD are treated like saints.

The reality is that they both are just charging what they calculate will maximize their returns. Neither is evil, and neither is your friend.
 
Stoly said:
I think the question could answer itself.

There's the GTX1660Ti vs RTX2060. It's too early on the stearsurvey to know which has a bigger share (the 2060 is higher on the list for now), but in a few months we'll have a better picture.
Click to expand...

Despite the similar names, they aren't in the same category. The 2060 is clearly faster than the 1660Ti. I'm talking about a 1:1 performing card minus RTX. So the question can't answer itself.
 
Stoly said:
Yeah, I'm sure stock holders hate that :rolleyes::rolleyes:
Click to expand...

I'm pro business and I get it. But at the same time when Company X prices Product Y at a perceived fucking ridiculous price, Customer Z can be turned off by this and will in turn NOT buy it.

X = Nvidia
Y = 2080Ti
Z = Me

They aren't getting another dime out of me until pricing returns to reality. And if it doesn't...I'll have more free time away from my computer, I'll turn down the eye candy, or just stick to a smaller monitor. I'm simply not going to pay over $1000 for a video card no matter how Nvidia thinks they should price it. Especially when I can build out a decently high end system minus the video card for under $1000 for all the rest of the parts combined.

At the same time, obviously, I can't speak for anyone else as how you spend your money is your own fucking business.
 
Last edited:
  • Like
Reactions: Nenu
like this
Snowdog said:
You do realize that even the 2060 die is very Large.
You do realize that NVida has a proportional amount of RT cores at every level, Titan RTX has double the Shaders as 2070 and it has Double the RT cores as well. If all the RT cores they needed was the amount in the 2060 they would have stopped there and added more Shaders.
Click to expand...
Turing SM is not new design, it is taken from Volta with some tweaks and RT core added in to it.
Adding more CUDA cores to SM block would most probably require a lot more effort, especially to do right.
Smaller RT core might not even be possible or could also require a lot more work.
Besides even if like I suspect there is an excess in RT performance compared to how it is being used (mostly shaders are taxed) it only means that game engines could be designed in such a way to take advantage of it, meaning they could use more ray tracing with simpler shaders.

But of course, you think know better than NVidia does when it comes to designing RT GPUs. :rolleyes:
Click to expand...
Where did this even come from?
I didn't say they should redesign their chip but that as it is there is no shortage of ray-intersection processing power to worry about.
 
XoR_ said:
Where did this even come from?
I didn't say they should redesign their chip but that as it is there is no shortage of ray-intersection processing power to worry about.
Click to expand...

If you are claiming 2060 levels of RT HW is all they need, then when they have double this on high end cards it would clearly be a waste, and the implication is that you think you know more about this problem space than NVidia.

Further beyond that we now have Q2 RTX, which is pure ray tracing, and is almost certainly limited by RT HW.

At 1080p, a 2060 squeaks 60 FPS with global illumination turned down to medium, and a 2080 Ti, nearly doubles it.

That isn't shader performance, it's RT HW performance.

That is with an ancient game with ultra low poly counts.

Full Ray tracing on a modern title would almost certainly crush even a Titan RTX.

You are going to see RT HW increase with each generation, to get closer to doing full RT on modern games.
 
Snowdog said:
If you are claiming 2060 levels of RT HW is all they need, then when they have double this on high end cards it would clearly be a waste, and the implication is that you think you know more about this problem space than NVidia.
Further beyond that we now have Q2 RTX, which is pure ray tracing, and is almost certainly limited by RT HW.
At 1080p, a 2060 squeaks 60 FPS with global illumination turned down to medium, and a 2080 Ti, nearly doubles it.
That isn't shader performance, it's RT HW performance.
That is with an ancient game with ultra low poly counts.
Full Ray tracing on a modern title would almost certainly crush even a Titan RTX.
You are going to see RT HW increase with each generation, to get closer to doing full RT on modern games.
Click to expand...
Where did I say 2060 RT shader performance was all card like 2080Ti needed? Eh?
I said RT core performance is more than enough suggesting that they do not need to increase RT core count or its size any furhter, at least for current revision of DXR that does not have any fixed shader program templates that can be made into ASIC to save shader performance by not needing to execute them on shaders.

And what is with your comprehension capabilities? Did I not explain each ray intersection is handled by shader program? You think "shader program" is executed on what? Quake2 RTX makes shaders on my RTX card scream just as strong as stroggs dying from my unbeatable blaster... hyper ofc ; )

If you still do not understand how this works go read some documentation or to save time tutorial like this one https://devblogs.nvidia.com/vulkan-raytracing/ and https://iorange.github.io/

Do your homework and all confusion will leave your mind. You will possess knowledge of how to properly shoot rays from your cameras and surfaces and recursively mix their output into brilliant vision of enlightened reality =)

ps. Reading about de-noising would probably be also good, it will help to get grasp how to recoustruct image from seemingly random noise.
 
XoR_ said:
Where did I say 2060 RT shader performance was all card like 2080Ti needed? Eh?
I said RT core performance is more than enough suggesting that they do not need to increase RT core count or its size any furhter, at least for current revision of DXR that does not have any fixed shader program templates that can be made into ASIC to save shader performance by not needing to execute them on shaders.

And what is with your comprehension capabilities? Did I not explain each ray intersection is handled by shader program? You think "shader program" is executed on what? Quake2 RTX make shaders scream on my RTX card just as strogs scream dying from my unbeatable blaster... hyper ofc ; )

If you still do not understand how this works go read some documentation or to save time tutorial like this one https://devblogs.nvidia.com/vulkan-raytracing/ and https://iorange.github.io/

Do your homework and all confusion will leave your mind and you will be enlightened knowing how to shoot rays from your cameras properly =)
Click to expand...

I said RT HW, not Shader performance.

I what kind of shader work do you think is being done in Q2-RTX??

It's going to be extremely minimal. It's unlikely in the extreme that Q2-RTX is shader limited.

If you are correct, 2060 is the most RT HW we need and NVidia will correct that going forward, by reducing RT HW proportion on high end models.

But I think you have completely wrong, and RT HW will increase across the board on future models.

Edit:

Here is a visual to point out why you are totally off the mark. It's time/load graph of a Metro Exodus frame.

Note that while the RT cores are busy doing their RT work, FP32 and Int32 cores are extremely lightly loaded. Negligibly so.

Ray Tracing is NOT shader limited.


geforce-rtx-gtx-dxr-one-metro-exodus-frame.png
 
Last edited:
Mode13 said:
Note: It's been some years since I've written any 3d rendering engine, and I did so as a hobby, so do your own research on anything I say.

The problem is in how raytracing works. Rather than baked in static lighting that we're used to with rasterization, ray tracing is essentially breaking a scene all the way down to the level of photons. Every single pixel you see is being traced from every light source to the object right down to the pixel, EVERY SINGLE FRAME. When you move in real life, the angle between the objects you're seeing and the light source (light bulbs, the sun, etc) is constantly changing causing a shift in perceived brightness and color. Multiply your screen resolution (1920x1080) and you'll be at ~2,000,000, consider how many rays are being cast per frame, and then multiply that by how many frames per second you expect to game to run, and you should now see the problem. That's just scratching the surface without thinking about how light interacts and scatters or moving and static objects casting shadows and reflections. Rasterized games don't have that problem because the lighting doesn't change as you move through a scene - or if it does it's happening in a pre-calculated (baked in) manner, rather than being computed in real time.

Hopefully that makes sense. There are many great technical references available online if you want to learn how it actually works.
Click to expand...

and then this fuggin guy just drops the best comment of 2019 possibly the best comment ive ever read on hardocp. i never "got" what this tray-dacing was all about and he just enlightens me like a child.
 
MangoSeed said:
Nvidia hardware implementation sucks compared to what? We literally have no other RT implementation to compare it against.
Click to expand...

Not quite true.
I have seen real time raytracing done in 3 ways:
RTX hardware - RT cores/Tensor-cores) on Turing cards doing DXR. (NVIDIA)
RT in shaders - Some Turing cards/Pascal cards (NVIDIA)
RT on CPU - Intel has done this for years as a tech concept (Intel)

One players is missing frrom the "party" though" (AMD)
 
purple_monster said:
and then this fuggin guy just drops the best comment of 2019 possibly the best comment ive ever read on hardocp. i never "got" what this tray-dacing was all about and he just enlightens me like a child.
Click to expand...

Not sure what the problem is. Look at the title of this thread and then you wonder why I sought a simplified way to explain ray tracing to people who might not understand it? Post wasn't even meant for you.

Apologies if I belittled anybody, but I think you're being impressively overdramatic.
 
Factum said:
Not quite true.
I have seen real time raytracing done in 3 ways:
RTX hardware - RT cores/Tensor-cores) on Turing cards doing DXR. (NVIDIA)
RT in shaders - Some Turing cards/Pascal cards (NVIDIA)
RT on CPU - Intel has done this for years as a tech concept (Intel)

One players is missing frrom the "party" though" (AMD)
Click to expand...

In an actual shipping game...
 
Mode13 said:
Not sure what the problem is
Click to expand...

You're being complimented ;)


[some people haven't drawn the line between the rasterization done in real-time consumer graphics and the type of rendering done for CGI, so when it dawns on them that ray tracing is what they've been staring at for years and that it's moving from being an offline render to a real-time render, they might get over-dramatic]
 
Mode13 said:
Not sure what the problem is. Look at the title of this thread and then you wonder why I sought a simplified way to explain ray tracing to people who might not understand it? Post wasn't even meant for you.

Apologies if I belittled anybody, but I think you're being impressively overdramatic.
Click to expand...

my comment was ambiguousand unclear when i reviewed it again. it was meant as a 100% compliment, sorry for being a weirdo.
 
Mode13 said:
Not sure what the problem is. Look at the title of this thread and then you wonder why I sought a simplified way to explain ray tracing to people who might not understand it? Post wasn't even meant for you.

Apologies if I belittled anybody, but I think you're being impressively overdramatic.
Click to expand...

The title is not the only place the OP is utterly and truely lost...his claims are even whackier that the title...
 
IdiotInCharge said:
You're being complimented ;)


[some people haven't drawn the line between the rasterization done in real-time consumer graphics and the type of rendering done for CGI, so when it dawns on them that ray tracing is what they've been staring at for years and that it's moving from being an offline render to a real-time render, they might get over-dramatic]
Click to expand...

I had always heard that Pixar ray traced their movies but now that the topic is in the limelight, I'm reading that Toy Story all the way up to Cars were entirely rasterized. I was curious to know how long a single frame from one of those movies would have taken to render in the 90s. When I was in school I recall doing a still image of a few transparent balls (or maybe it was some simple fractal) inside of a cube with one light source, I remember that taking many hours to render a single frame on a 486 or early Pentium. It would be mind-blowing if a scene as detailed as a modern pixar movie (or AAA game) could be rendered in real time now.


purple_monster said:
my comment was ambiguousand unclear when i reviewed it again. it was meant as a 100% compliment, sorry for being a weirdo.
Click to expand...

Ah my fault, guess I missed half the thread and lost the context. Every year that goes by sees me more and more out of touch with what people are saying anymore, especially online... I blame the anxiety.. X_X
 
Mode13 said:
I had always heard that Pixar ray traced their movies but now that the topic is in the limelight, I'm reading that Toy Story all the way up to Cars were entirely rasterized. I was curious to know how long a single frame from one of those movies would have taken to render in the 90s. When I was in school I recall doing a still image of a few transparent balls (or maybe it was some simple fractal) inside of a cube with one light source, I remember that taking many hours to render a single frame on a 486 or early Pentium. It would be mind-blowing if a scene as detailed as a modern pixar movie (or AAA game) could be rendered in real time now.




Ah my fault, guess I missed half the thread and lost the context. Every year that goes by sees me more and more out of touch with what people are saying anymore, especially online... I blame the anxiety.. X_X
Click to expand...


https://en.wikipedia.org/wiki/Toy_Story

Under Galyn Susman and Sharon Calahan, the lighting team orchestrated the final lighting of the shot after animation and shading. Each completed shot then went into rendering on a "render farm" of 117 Sun Microsystems computers that ran 24 hours a day.[36] Finished animation emerged in a steady drip of around three minutes a week.[62] Depending on its complexity, each frame took from 45 minutes up to 30 hours to render. The film required 800,000 machine hours and 114,240 frames of animation in total.[37][57][63] There are over 77 minutes of animation spread across 1,561 shots.[59] A camera team, aided by David DiFrancesco, recorded the frames onto film stock. To fit a 1.85:1 aspect ratio, Toy Story was rendered at a mere 1,536 by 922 pixels, with each of them corresponding to roughly a quarter-inch of screen area on a typical cinema screen
Click to expand...
 
Sidenote:

Seems NVIDIA's solution is perfectly fine for PIXAR:
https://nvidianews.nvidia.com/news/...illions-of-3d-artists-and-designers-this-year


“Pixar artists already rely on NVIDIA ray tracing, and RTX more than doubles the performance they will see. We’re excited to use RTX on our upcoming films,” said Steve May, CTO of Pixar Animation Studios.

“RTX technology is a game changer for our architectural visualization pipeline,” said Gamma Basra, partner and head of Visualization at Fosters+Partners. “We can iterate options to ascertain the optimal design in real time without the need to wait hours for the render to come back.”
Click to expand...

The OP really misfired here...

/que To MUCH Ray tracing next...isn't that the usual mantra?
 
Aireoth said:
Unity and Unreal?!! o_O What engines do you think developers use en mass these day? Infinity?
Click to expand...

Ok so Unity is geared towards people starting out and indies. I'm not sure that's where we are going to get anything too impressive visually for at least a while

For UE4, sure it's popular, but I don't find myself playing anything that uses it. Lots of second and third tier titles in the list of games that UE4 is used by.

Hopyfully is doesn't turn out to be like async compute
 
blackmomba said:
Ok so Unity is geared towards people starting out and indies. I'm not sure that's where we are going to get anything too impressive visually for at least a while

For UE4, sure it's popular, but I don't find myself playing anything that uses it. Lots of second and third tier titles in the list of games that UE4 is used by.

Hopyfully is doesn't turn out to be like async compute
Click to expand...

So you where wrong.
 
blackmomba said:
Ok so Unity is geared towards people starting out and indies. I'm not sure that's where we are going to get anything too impressive visually for at least a while

For UE4, sure it's popular, but I don't find myself playing anything that uses it. Lots of second and third tier titles in the list of games that UE4 is used by.

Hopyfully is doesn't turn out to be like async compute
Click to expand...
How on earth did you come to the notion to compare RT with Ascyn compute?
 
Factum said:
Then there are two options...happy now?
Click to expand...

Thank you for being so concerned about my happiness :) Also for agreeing with me that there are no comparable hardware implementations of RT to support the OP.
 
MangoSeed said:
Thank you for being so concerned about my happiness :) Also for agreeing with me that there are no comparable hardware implementations of RT to support the OP.
Click to expand...

Don't put words into my mouth...the OP is decoupled from reality...like your trying to attribute false claims to me.
Funny how RT makes people turn their brains off.
 
Snowdog said:
I said RT HW, not Shader performance.
I what kind of shader work do you think is being done in Q2-RTX??
It's going to be extremely minimal. It's unlikely in the extreme that Q2-RTX is shader limited.
If you are correct, 2060 is the most RT HW we need and NVidia will correct that going forward, by reducing RT HW proportion on high end models.
But I think you have completely wrong, and RT HW will increase across the board on future models.
Edit:
Here is a visual to point out why you are totally off the mark. It's time/load graph of a Metro Exodus frame.
Note that while the RT cores are busy doing their RT work, FP32 and Int32 cores are extremely lightly loaded. Negligibly so.
Ray Tracing is NOT shader limited.
Click to expand...
Where did I mention RT performance of RTX 2060 being most that we will need?
Or where did I say Nvidia will not increase "RT hardware"?
Where do you pull these things from?

As for graph, did it occurred to you that shader workload during tracing rays part is not done at the same time than actual pixel color calculations? These cards have gigabytes of memory, they can remember state of programs and calculated values. Fully programmable means also that power of choosing what is done when is in hands of developers.
And what is with RT cores working for like 16% of frame rendering time? Is it not proof of what I way saying all along that RT hardware is already overpowered?
Besides, obviously if ray tracing effects took only as much time as RT cores and Tensor cores are working then it wouldn't impact final frame time so much as it does, now would it? =)
Same chart but completely opposite interpretation.

There are reasons RTX cards are designed like they are and some reasons have purpose. Even if this purpose is afterthought and dictated by other restraints it most probably influenced decision process. One unimaginative reason I can give you is: to not allow competition to beat them in their own game, at least not easily. Other reason would be to make RT engine purposefully powerful enough to make it possible for unexperienced developers who have no idea what they are doing to have at least somewhat playable frame-rates. With less RT power butchered implementations like these would run even slower... But like I said, this was probably not on purpose as much as result of hardware design constraints and they could not make RT cores any smaller or put less of them or put more CUDA cores. Nvidia will definitely concentrate efforts on powering up is Tensor cores. Machine learning is where the real money are. RT stuff itself is far less important for Nvidia and I bet they made it to have as small impact on die size as possible.

Texture mapping units and other units are also being used by RT cores and there is all the memory access issues to consider when pondering about this design and why it is as it is. There will be definitely a ton of tweaking to make it faster, this is for sure. If they will even bother increasing CUDA cores per SM block I am not really sure but as it is hardware DXR seems to not be limited by RT cores performance and while making it more powerful would help (as your chart suggest) it would not by by much (as your chart also clearly suggests). At this time performance of DXR effects seems to be in hands of developers... as is implementing DXR at all.

Like I said, if you are truly interested in knowing how this all works go read some DXR documentation. I am enlightened in art of extromission so you won't win any discussion with me until you get hold of rays of light yourself ~_^
 
XoR_ said:
Where did I mention RT performance of RTX 2060 being most that we will need?

...

Is it not proof of what I way saying all along that RT hardware is already overpowered?
Click to expand...

Maybe it would help if you stopped contradicting yourself even within the same post. Already overpowered = The most that we will need.

No, it's not an indication that RT HW is already overpowered. This is a Hybrid title, that are mostly Raster Based with some RT effects added on, so the time they spend doing RT calls would not be as high a proportion as they would in a full RT title. We are forced to scale back RT effects at this time because they HW is under-powered.

But it is an indication that you were wrong about the shader impact of doing intersections.

During the actual ray tracking, it's the intersections that are the main body of work. It's trivial applying color to the pixel once you know the intersection. Here is a blow up of that frame. While the intersections are being calculated (big green area) you can see a VERY small amount of shader work is happening in conjunction (grey and purple under the green), this is the trivial work of adding color info from each intersection.

Now if this were something like a Q2-RTX frame, it would certainly have proportionally more intersection work for RT cores and less shader work, because Q2-RTX is pure ray tracing, not a hybrid title.
geforce-rtx-gtx-dxr-metro-exodus-rtx-rt-core-dlss-frame-expanded.png
 
Status
Not open for further replies.
Back
Top