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

[leadeater]

3 hours ago, Settlerteo said:

I see your point, but just tell me something. If all of us stopped buying any cards over 300-350$(which was the enthusiast price range in the not so distant past), would they drop the prices or do you think they would keep up. Marketing tricks always aim for the increase on sales but sometimes they also aim at higher prices(they're just selling marketing along with the card itself).

Well you would have to sacrifice performance to get a lower price. While I do hate how high the prices are for the top end hardware it's not going to get cheaper simply by people not buying them, cost to develop and manufacture does not go down through lack of sales.

 

If the manufacturing cost is not too high you could get lower prices but that would have to mean time between architectures would increase significantly or performance increases drop, R&D expenses would have to go down as that is where the price increases are in large part coming from.

 

 

https://ycharts.com/companies/NVDA/r_and_d_expense

 

Nvidia is spending more than AMD who covers both CPUs and GPUs.

 

AMD for reference

https://ycharts.com/companies/AMD/r_and_d_expense

[Brooksie359]

4 hours ago, Settlerteo said:

I see your point, but just tell me something. If all of us stopped buying any cards over 300-350$(which was the enthusiast price range in the not so distant past), would they drop the prices or do you think they would keep up. Marketing tricks always aim for the increase on sales but sometimes they also aim at higher prices(they're just selling marketing along with the card itself).

The top of the line from nvidia has been over 350 for a long time now. You can see their top of the line cards being over 350 going way back so I am unsure what you are talking about. 

[Tech Enthusiast]

People keep demanding new stuff. 

And people expect wonders every 1-2 years.

 

Also people: Not willing to pay more for all the stuff, expecting prices to stay the same.

 

That would work, if we would wait 5-8 years between generations, like consoles! Do we want console like performance for 8 years? No? Well, guess we will have to pay more then.

[Brooksie359]

3 minutes ago, Tech Enthusiast said:

People keep demanding new stuff. 

And people expect wonders every 1-2 years.

 

Also people: Not willing to pay more for all the stuff, expecting prices to stay the same.

 

That would work, if we would wait 5-8 years between generations, like consoles! Do we want console like performance for 8 years? No? Well, guess we will have to pay more then.

If it cost more to make then it will cost more to buy. It's a pretty simple concept. If it cost the same as last gen to make then they could sell it for the same price as last gen for each tier but that is not the case. To be fair you still get slightly more performance for your money and new features. 

[CarlBar]

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

DXR is not the gimmick, you can call RTX an nvidia gimmick if you want but raytracing is coming to all platforms one way or another. DXR makes that reality

Was replying about this

I should point out that DICE also mentions wanting to add RT Core support as soon as possible but we dont know if that will happen. As far as them using Tensor cores, the way they say it in the video was it was developed using the methods available on Titan V's which was tensor core only. As some of you will remember, ray tracing in games wasn't really even talked about until the star wars demo, courtesy of tensor cores. You can't do real-time RT on a pascal card because no tensor cores and no game is implementing cpu only raytracing yet. (Ok I think there's one with like super shit quality RT)

 

I'm about to go back and watch the linked video but the breakdown someone did of their demo build running on the Titan V explicitly noted that whilst the card had Tensor cores, they were not being used.

 

EDIT: And it's the same video that mentions this, 9:20 ish for a timestamp.

[crystal6tak]

I'm getting a bit confused about the terminology used here, are RT cores and tensor cores different things?

 

Is it correct to say Titan V has shader (cuda) + tensor cores; 2080 ti has shader + tensor + RT cores; with traditional cards (e.g. 1080 ti) only having shader cores?

[Tech Enthusiast]

1 hour ago, crystal6tak said:

I'm getting a bit confused about the terminology used here, are RT cores and tensor cores different things?

 

Is it correct to say Titan V has shader (cuda) + tensor cores; 2080 ti has shader + tensor + RT cores; with traditional cards (e.g. 1080 ti) only having shader cores?

exactly that, yes.

[S w a t s o n]

2 hours ago, CarlBar said:

 

I'm about to go back and watch the linked video but the breakdown someone did of their demo build running on the Titan V explicitly noted that whilst the card had Tensor cores, they were not being used.

 

EDIT: And it's the same video that mentions this, 9:20 ish for a timestamp.

For denoising which is the intended use for Tensor cores in a proper RTX implementation. RTX only comes in two flavours: Tensor core raycasting (Star Wars Demo, Battlefield V) or RT Core raycasting with Tensor Core denoise (Metro Exodus I hope). There is no RTX implementation currently that will run with without at least one of those technologies.

[Sauron]

Dedicated RX cores are probably useful, but the 2080ti doesn't have enough of them to leave the titan v in the dust. Remember, the largest part of the card is doing almost nothing when using RX on the dedicated cores. Hopefully next generation we'll have cards with more RX cores, as well as the regular non-RX units.

[S w a t s o n]

1 minute ago, Sauron said:

Dedicated RX cores are probably useful, but the 2080ti doesn't have enough of them to leave the titan v in the dust. Remember, the largest part of the card is doing almost nothing when using RX on the dedicated cores. Hopefully next generation we'll have cards with more RX cores, as well as the regular non-RX units.

It's not using the RT cores, that's why Titan V is nearly equal, they are both operating on tensor core method of RTX for BFV, expect to see a difference in other games

[mr moose]

12 hours ago, Settlerteo said:

I see your point, but just tell me something. If all of us stopped buying any cards over 300-350$(which was the enthusiast price range in the not so distant past), would they drop the prices or do you think they would keep up. Marketing tricks always aim for the increase on sales but sometimes they also aim at higher prices(they're just selling marketing along with the card itself).

 

The others answered this question well enough, but I'll just that Consumers almost never have enough impact to change market forces.  So the likelihood of people not buying the top end is almost none existent.     Fir example one of the largest consumer boycotts we have ever witnessed was against nestle and started in the 80's, I have met people in the late 90's who still promote that boycott, yet to be honest I don't think it's effects are even measurable in Nestle's business activities.

[bcredeur97]

so how does one enable the ray tracing setting in bf 5 with an "incompatible" card?

[bcredeur97]

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

Polyphony Digital (makers of Gran Turismo) are developing their own RT engine, so hopefully Navi has that rumoured AI silicon or the PS5 has so many zen cores they are just gonna brute force it. Imagination Technologies and other chip makers are all looking into this. Raytracing in mobile games will be a thing. The whole industry is ready.

As CPU core counts take off in the next 5 years I expect to see some CPU raytracing implemented

what you say reminds me of physX lol

[leadeater]

20 minutes ago, bcredeur97 said:

what you say reminds me of physX lol

Well at least RTX doesn't require a second GPU or a special RTX accelerator card, something can almost always be worse or get worse lol

[CarlBar]

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

For denoising which is the intended use for Tensor cores in a proper RTX implementation. RTX only comes in two flavours: Tensor core raycasting (Star Wars Demo, Battlefield V) or RT Core raycasting with Tensor Core denoise (Metro Exodus I hope). There is no RTX implementation currently that will run with without at least one of those technologies.

 

The video doesn't say anywhere they're doing the ray tracing on the Tensor Cores ethier.

[S w a t s o n]

2 minutes ago, CarlBar said:

 

The video doesn't say anywhere they're doing the ray tracing on the Tensor Cores ethier.

They clearly state the raytracing was developed using the technologies available on Titan V. That's tensor cores and tensor cores alone. RTX doesn't exist without tensor cores, you can check nvidia's various whitepapers and other information to confirm.

This video is just the best source we have directly from DICE, you're proposing something that doesn't exist anywhere at all. Nvidia showed pascal frametimes trying to run RTX at Turing launch, it doesn't work.

[CarlBar]

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

They clearly state the raytracing was developed using the technologies available on Titan V. That's tensor cores and tensor cores alone. RTX doesn't exist without tensor cores, you can check nvidia's various whitepapers and other information to confirm.

This video is just the best source we have directly from DICE, you're proposing something that doesn't exist anywhere at all. Nvidia showed pascal frametimes trying to run RTX at Turing launch, it doesn't work.

 

And they also clearly state they aren't using the tensor cores for denoising, which is all NVIDIA themselves have claimed they can be used for. There is zero evidence, (that i've seen), beyond supposition that they're not running everything in the shader pipeline.Which i admit is a bit shocking.

 

But if where going to lay battle of supposition: If Tensor cores can do ray tracing well enough to be this functional why didn't NVIDIA just tack more Tensor cores on rather than developing RT cores. They can clearly get playable frame rates without them so why develop whole new hardware they don't need?

 

Personally i suspect the almost completely nonexistent performance of the half precision performance explains a lot of the gap between the 20 and 10 series GPU's in this respect. I also suspect the fact the DICE per their own admission is running a seriously cut down version of even the reflection RTX implementation, (when the cards were supposed to be able to do shadows, reflections and lighting all at once), plays an enormous partm, they're also having issues with their denoiser not working well in some scenarios which makes me suspect it's, (naturally), not as high quality s Tensor cores could allow which obviously cuts down the processing overhead.

 

Basically it looks like DICE cut a ton of corners to pull it off whilst taking advantage of improvements in the standard shader pipeline.

[Trik'Stari]

I can't speak about graphics, as they looked fined to me on BFV with a 1080ti water cooled rendering at 4k ultra/high.

 

What I can say, is that the sound was ABSOLUTE GARBAGE. This from someone who loves the Battlefield franchise for it's immersive and often quite useful sound design.

 

But in BFV? It was worse than any CoD game I've ever played, even set to War Tapes.

[S w a t s o n]

3 hours ago, CarlBar said:

And they also clearly state they aren't using the tensor cores for denoising, which is all NVIDIA themselves have claimed they can be used for. There is zero evidence, (that i've seen), beyond supposition that they're not running everything in the shader pipeline.Which i admit is a bit shocking.

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?
 

[S w a t s o n]

@CarlBar I understand it's not literally spelled out but if you look at all the numbers and information it's pretty clear.

See my post earlier in the thread comparing how the core handles ray tracing on pascal vs turing. Pascal has to do thousands of instructions PER RAY just to calculate if it even intersected with something. It's horrifically inefficient. Tensor can infer the rays and also denoise but it's being pushed for denoise because it's more suited to that task hence RT Cores. Do you think the volta cards raytrace so much faster than pascal for no reason?

[leadeater]

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

Tensor can infer the rays and also denoise but it's being pushed for denoise because it's more suited to that task hence RT Cores. Do you think the volta cards raytrace so much faster than pascal for no reason?

Or from the information in the picture that Turing is supposed to be 4 times faster at ray casting than Volta, at least twice as much since that Star Wars demo was ray casting and denoising on the same set of Tensor cores.

[Crunchy Dragon]

22 hours ago, Tech Enthusiast said:

The interesting part is not that it runs on a Titan V. 

Since DXR is a clean DX12 thing, no one that knows how this works was expecting ANY GPU not to "run" it.

 

The point is that a Titan V runs it well. Very well actually.

Granted: The same User confirmed less impressive results on other maps, but we are still talking about playable framerates, which where sold as not being possible without RT cores. And honestly, since RT has been done for a good while now, we never saw any "playable" implementation up to now. I mean, it is not like people doing this for work did not try to make it go faster or something. 

 

It seems like Tensor Cores are indeed very capable of RT work. The real question is: How capable are they really? Could RTX go without RT by having more Tensor?

More Tensor would make it vastly superior in other tasks as well,...

RTX card benchmarks were still playable framerates.

 

Unplayable is 30 and below, even as low as 50 is still good. This is why I still believe that people should stop complaining about "muh framerates", it's not like it runs at 30 fps to let ray tracing happen; it's still perfectly playable.

 

It's still worth noting that RTX will only get better in games. This isn't the final version that gamers will be stuck with forever.

[pas008]

Considering the titan has like 800 more cuda 

Brute forcing higher fps?

[zhubaohi]

On 1/2/2019 at 2:49 AM, sazrocks said:

How much do we know about the architectures of the RTX cores and Tensor cores? Is it possible that they are similar?

RTX 2080ti performs very similar to Titan V in deep learning territory, and they all perform very similar in ray tracing application. That's probably a sign that ray tracing and deep learning uses a similar technology.

[S w a t s o n]

6 hours ago, zhubaohi said:

RTX 2080ti performs very similar to Titan V in deep learning territory, and they all perform very similar in ray tracing application. That's probably a sign that ray tracing and deep learning uses a similar technology.

Tensor cores are matrix multiply units.

https://www.nvidia.com/en-us/data-center/tensorcore/

 

2080ti and Titan V share similar number of tensor cores hence similar AI performance. In this case of RTX they are using the tensor cores, hence similar ray tracing performance. 2080ti and titan v will not have comparable raytracing performance in later RTX games that use RT Cores.

 

RT Core:

 

Basically, it doesn't seem like the RT Core is similar to the Tensor Core. The matrix multiply units (Tensor) are used to build a deep learning model for the game so that it can basically infer the correct rays to trace rather than calculate it. It's doing the same thing it would be doing for an AI research project but raytracing instead of say object recognition in a live video.

RT Core just sits there and does the math to traverse the BVH and KNOW the correct rays rather than sort of guessing. Also each SM on Turing has 1 RT Core but 8 Tensor Cores and I think RT Cores take up nearly same die area meaning each RT Core is huge compared to tensor