Nvidia Turing Announced

[Froody129]

What I would imagine Nvidia doing is making a "consumer" Titan card, except now it's not just overpriced but includes some of this extra hardware. This would give some credibility to the 'Titan' name and differentiate it from the 80ti card. 

[AlfaProto]

Looking at RX Navi, it looks like Nvidia pulls way ahead once again, leaving RX Navi into the dust. Looks the the RX series is a total flop.

[S w a t s o n]

28 minutes ago, Froody129 said:

What I would imagine Nvidia doing is making a "consumer" Titan card, except now it's not just overpriced but includes some of this extra hardware. This would give some credibility to the 'Titan' name and differentiate it from the 80ti card. 

That's what I expected at first but at this point the rumours point to RTX 2080 so it should have RT and Tensor cores, unless Nvidia goes full retard, and cuts down and re-spins volta for GDDR6

(not really rumors now, look at dis)

 

(Unrelated but I have an exam tomorrow and a final essay due and this whole thing delayed me starting by like 4 hours but I got dat essay done, gonna study overnight, WORTH IT)

[leadeater]

9 hours ago, Swatson said:

 

Equivalent render farm on turing

For interest sake the Nvidia DGX-2 with 16 Tesla V100's uses 10KW of power. 32 total RTX GPUs at 13KW with the potential to run target workloads at 4 times performance efficiency, damn.

[Bananasplit_00]

That's a good looking quadro tho. They tend to look like shit but the one in the picture looks pretty nice imo

[S w a t s o n]

3 minutes ago, Bananasplit_00 said:

That's a good looking quadro tho. They tend to look like shit but the one in the picture looks pretty nice imo

The shroud is pretty reflective and Jensen literally spent a solid minute memeing about it and shining it into the camera because muh ray tracing

[S w a t s o n]

So Nvidia is claiming 10 giga rays (Giga denotes billion so 10 Billion rays) per second but even at 8k, 5 rays per pixel is only 165 million rays. I think they might be counting shadow rays and bounces? Still that's more than an order of magnitude difference....

 

Edit: This is why I shouldn't post at 6am while studying since the night before. 165 million PER FRAME at 22FPS (approx star wars demo FPS) is 3.6 Giga Rays not including shadow rays or bounces so it's within reason i suppose.

[kilgore_T]

I want dem gigarayz

[SADS]

1.21 Gigarayz?!

https://www.youtube.com/watch?v=mjCRUvX2D0E

[S w a t s o n]

Just now, Sampsy said:

I imagine not. ~20fps isn't really playable and the consumer GPUs would have smaller dies and wouldn't even be able to achieve that. It's exciting as it looks like ray tracing in games may be a real thing in the near future, but it's not quite ready yet. 

the 22FPS achieved in star wars demo should not be taken to be what a game would run at, the RTX implementation will not be as nice. You can see a preview of game quality RTX if you look at video for Metro Exodus

[Age of 9394]

It the new gpu come with new logo because all the Nvidia GeForce Youtube channels logo became green circle.

[laminutederire]

4 hours ago, AlfaProto said:

Looking at RX Navi, it looks like Nvidia pulls way ahead once again, leaving RX Navi into the dust. Looks the the RX series is a total flop.

Well Nvidia creates a problem and solves it itself. When is that impressive?

I'm pretty sure their approach cannot be extended to mainstream for so many reasons. It leverages technologies which aren't easily flexible. The only way it can work is that if it's a framework like model where we all get screwed because since it's closed,AMD or Intel won't be able to properly optimize for it and Nvidia can get away with poor optimization themselves since they have to fight minimal optimization. And we get something useless and almost unusable..

[Amazonsucks]

5 hours ago, Froody129 said:

What I would imagine Nvidia doing is making a "consumer" Titan card, except now it's not just overpriced but includes some of this extra hardware. This would give some credibility to the 'Titan' name and differentiate it from the 80ti card. 

Thats what the first Titan was: GK110 with full DP enqbled but no ECC at a fraction of the price of the K20 or K6000. It was only with Maxwell and the advent of the 200(became 102) medium chip that the Titan and x80ti became essentially the same card.

 

But thats because the 102 chip is just a bigger 104 and not at all like the GP and GV100 chips which have DP performance, ECC registers and RAM like the old GK110. People forget that the only time since Kepler that Nvidia has put their big chip in a consumer card is the $3000 Titan V. Maxwell didnt even have a real big chip.

[Amazonsucks]

55 minutes ago, laminutederire said:

Well Nvidia creates a problem and solves it itself. When is that impressive?

I'm pretty sure their approach cannot be extended to mainstream for so many reasons. It leverages technologies which aren't easily flexible. The only way it can work is that if it's a framework like model where we all get screwed because since it's closed,AMD or Intel won't be able to properly optimize for it and Nvidia can get away with poor optimization themselves since they have to fight minimal optimization. And we get something useless and almost unusable..

Really? The thing that computer graphics people have tried to do for decades is finally here and its useless and unusable?

 

Let AMD make their own real time ray tracing to compete. Thats what they should do if they want to become relevant in graphics.

[mynameisjuan]

1 hour ago, Sampsy said:

~20fps isn't really playable

20fps is playable, shit experience but playable. 

 

These cards are pretty excited 

[williamcll]

now I'm interested in those Navi cards.

[TacoSenpai]

But can it run Crysis? 

 

Spoiler

Or you know at least a phone microwave time machine or something like that?

 

[Morgan Everett]

3 minutes ago, LordOTaco said:

But can it run Crysis? 

 

  Reveal hidden contents

Or you know at least a phone microwave time machine or something like that?

 

Yes, as well as tired memes about Crysis.

[S w a t s o n]

This is live right now but I assume will be available as VoD later. Cool little tutorial on programming raytracing on DXR which is what RTX is running on.

[PopReference]

I just had a maddening rumor worthy thought:

RTX 2080/70( ti) series with ray tracing capable cores

AND

GTX 1180/70/60.. ( ti) series without ray tracing cores

 

ALL the rumors are TRUE!!

[laminutederire]

4 hours ago, Amazonsucks said:

Really? The thing that computer graphics people have tried to do for decades is finally here and its useless and unusable?

 

Let AMD make their own real time ray tracing to compete. Thats what they should do if they want to become relevant in graphics.

What Nvidia does has been done since 2012... but with less quality. There is a reason why Disney isn't going to buy in chunk those cards. The ray tracing it does is minimal quality wise and takes a gigantic amount of power. At best it's an immature technique. Reason being ray tracing is just plain bad quality wise. There s not much you can really do with it. Path tracing is where the quality is at. And that's quite harder.

Then they propose and so denoiser? Cool. They do it at Disney since quite a long time now, and they still haven't gotten around the biggest hurdle that it will face in Nvidias case: it doesn't generalize well to different noise sources.

[S w a t s o n]

8 minutes ago, laminutederire said:

What Nvidia does has been done since 2012... but with less quality. There is a reason why Disney isn't going to buy in chunk those cards. The ray tracing it does is minimal quality wise and takes a gigantic amount of power. At best it's an immature technique. Reason being ray tracing is just plain bad quality wise. There s not much you can really do with it. Path tracing is where the quality is at. And that's quite harder.

Then they propose and so denoiser? Cool. They do it at Disney since quite a long time now, and they still haven't gotten around the biggest hurdle that it will face in Nvidias case: it doesn't generalize well to different noise sources.

The denoising is done with tensor cores, which have only existed since Volta, so unless disney was using volta you can't make that assumption. Even if they were using volta it's only been a couple years, and RT cores (which did not exist before) are the game changer now.

 

BTW pretty sure Nvidia knows more about graphics than disney. Disney's got money and talent but they buy their GPUs from AMD/Nvidia. It's like implying that a random car owner knows how to fix their car better than a dealership mechanic (assuming you dont get a dumb one)

[laminutederire]

Just now, Swatson said:

The denoising is done with tensor cores, which have only existed since Volta, so unless disney was using volta you can't make that assumption. Even if they were using volta it's only been a couple years, and RT cores (which did not exist before) are the game changer now.

Hum. It was done on CPUs and normal cuda cores and isn't the exact denoiser of Nvidia. It's one more advanced denoiser ready for production path tracers. They were testing it on Maxwell if I remember right. 

It's a game changer for Nvidias silky implementation. It's not going to revolutionise people using real path tracers which make more sense with a cpu architecture for many reasons.

The trend has been simple gpu path tracing is possible and faster in simple cases. The more you add complexity to effects and geometry the more they struggle because the memory layout isn't suited for it. (A lot of non contiguous memory call that can't really be batch processed efficiently unless you degrade the randomness and therefore the solution quality.)

[S w a t s o n]

1 minute ago, laminutederire said:

Hum. It was done on CPUs and normal cuda cores and isn't the exact denoiser of Nvidia. It's one more advanced denoiser ready for production path tracers. They were testing it on Maxwell if I remember right. 

It's a game changer for Nvidias silky implementation. It's not going to revolutionise people using real path tracers which make more sense with a cpu architecture for many reasons.

The trend has been simple gpu path tracing is possible and faster in simple cases. The more you add complexity to effects and geometry the more they struggle because the memory layout isn't suited for it. (A lot of non contiguous memory call that can't really be batch processed efficiently unless you degrade the randomness and therefore the solution quality.)

The point of using tensor cores is for AI assisted denoising though, AI seems to come up with novel and barely understandable solutions to problems. As in the computer engineers literally dont know exactly what is being done and could not replicate it manually.

[Underwrought]

15 hours ago, Swatson said:

My condolences to anyone that bought 1080ti, Titan xP or Titan V (at least you get slow RTX on Titan V) in the last 6 months

  Reveal hidden contents

 

As someone who just bought a 1080 ti a few days ago, im not worried about it at all.  It will take years for games to start to implement it fully, and years more for it to be refined further and become cost effective.  On top of all that I am willing to bet the performance increases are not what they are being hyped up to be.  But I guess we will see, I am obviously going to wait for the Ti version of this next gen anyway.  Or if I am wrong I might hopefully be able to evga step up?