Nvidia Ampere Discussion [2020-05-14]

[Jawed]

It doesn't really look like an Ampere vs Turing advantage so much as it's taking advantage of the huge memory bandwidth available on the 3080/3090 with RT on. It's allowing GA102 to stretch its legs more vs GA104 than in most titles, which does make sense given that RT enabled at 4k is about as bandwidth-heavy of a scenario as you're going to find.

It seems unlikely that consumer cards will get substantially more than 1TB/s bandwidth within a few years. Does that meant that RT performance is going nowhere?

On the other hand, perhaps NVidia will implement something like Infinity Cache, but say 256MB?

 

[arandomguy]

It seems unlikely that consumer cards will get substantially more than 1TB/s bandwidth within a few years. Does that meant that RT performance is going nowhere?

On the other hand, perhaps NVidia will implement something like Infinity Cache, but say 256MB?

RTX 3070 is also slightly faster than 2080ti in this workload which suggests that raw memory bandwidth isn't the only differentiator or limiter.

Infinity Cache is what AMD's chosen this generation (and then made to sure strongly market) but it may not be the only option to improve effective bandwidth (or utilization).

Cache sizes could increase, as cache sizes in Nvidia's case haven't really changed since Maxwell.

 

[Lurkmass]

No Monte Carlo rendering for you.

I guess the big difference is how is machine learning going to be helpful for solving multidimensional integrals like we see in rendering ?

With Monte Carlo methods, we start by sampling from a known model so having data is not a necessity here. With neural networks, we have an unknown model that we have to generate from the given datasets. In the first method we are aiming to simulate our model while in the second method we are training our model so there's is a profound distinction between the two.

If we always assume that the model we are given is going to produce more accurate and consistent results compared to our trained model then theoretically the first approach is going to have a built-in quality advantage ...

 

[Deleted member 13524]

A 3090 is 1.8x faster than the 2080TI. Normalized to die size Ampere delivers twice the performance.

Why would anyone compare performance-per-mm^2 between chips fabbed on distinct node generations?
The RTX 3090 / GA102 is also just 5% larger than the 980 Ti / GM100.

 

[CarstenS]

Why would anyone compare performance-per-mm^2 between chips fabbed on distinct node generations?

Marketing/PR departments really like x-factors of 2 and up. So much actually, that you see this kind of comparison even in some whitepapers with all kinds of companies.
As a consumer though, I wouldn't care.

 

[nAo]

I guess the big difference is how is machine learning going to be helpful for solving multidimensional integrals like we see in rendering ?

With Monte Carlo methods, we start by sampling from a known model so having data is not a necessity here. With neural networks, we have an unknown model that we have to generate from the given datasets. In the first method we are aiming to simulate our model while in the second method we are training our model so there's is a profound distinction between the two.

If we always assume that the model we are given is going to produce more accurate and consistent results compared to our trained model then theoretically the first approach is going to have a built-in quality advantage ...

There is a misconception on this forum that using ML/DL for graphics means throwing away any prior knowledge about a problem and just replace an old school solution with a black box that can learn its own degrees of freedom (e.g. network weights and biases). One can do that (and there is plenty of published work that does..) but it's a method that leads to brittle solutions that poorly generalize outside the training dataset. On the other hand there is also serious progress being made by using these new tools to augment classic solutions.

For instance importance sampling techniques are routinely used to lower the variance of Monte Carlo estimators (i.e. less noise) applied to rendering integrals, and recently researchers have showed how to further enhance these methods by using DL to learn the underlying probability distribution used to sample the integrand, leading to further reduction in noise (see below):

"Neural Importance Sampling" - https://arxiv.org/abs/1808.03856

And more sophisticated follow up work:

"Neural Control Variates: - https://arxiv.org/abs/2006.01524

These are early days and who knows what rendering will really encompass 10 years from now, but I have little doubt that DL is going to have a profound impact on how we generate images.

 

[Freemantle]

Let me know when AMD will release their $1000+ product. There's not a single one in their currently announced product portfolio.

Will do, not long now!

 

[troyan]

It does use RT cores.

Hardware RT is off per default. It must be activated per console command.

Am I missing something here? The 2080Ti and the 3070 trade blows in most games in all of the reviews - this doesn't appear to be a particular outlier performance-wise. In this case the 2080Ti is still faster than the 3070 with RT off, and a bit slower with RT on, but still in the same performance class.

It doesn't really look like an Ampere vs Turing advantage so much as it's taking advantage of the huge memory bandwidth available on the 3080/3090 with RT on. It's allowing GA102 to stretch its legs more vs GA104 than in most titles, which does make sense given that RT enabled at 4k is about as bandwidth-heavy of a scenario as you're going to find.

The 3090 is 80% faster and has only 50% more bandwidth than the 2080TI.

 

[DavidGraham]

Hardware RT is off per default. It must be activated per console command.

Are you sure about this info? Digital Foundry tested the game and found RT is accelerated by hardware on RTX GPUs if you selected Can it run Crysis?

 

[Kaotik]

Hardware RT is off per default. It must be activated per console command.

Source for this? It sounds really odd that developer would advertise it uses RT cores if you'd need to enable it via console

 

[Dictator]

Are you sure about this info? Digital Foundry tested the game and found RT is accelerated by hardware on RTX GPUs if you selected Can it run Crysis?

It sure is on by default. It detects the GPU on start.

 

[eastmen]

I ended up getting a zotac 3080 amp halo for $806 on their website with shipping and everything included. I felt this was a fine price as amazon and newegg would charge taxes so i would have paid similar anyway

 

[Deleted member 2197]

Regarding manual OC's and warranties, Nvidia still allows you to use any tool without breaking warranty. AMD limits warranty coverage to only RageMode and using any other method voids warranty.

This is an important point, because manual overclocking, including in the Wattman, is officially a violation of the warranty terms for AMD graphics cards. The Rage Mode, on the other hand, can be used to your heart's content - but is missing on the Radeon RX 6800 (Non-XT). The latter also allows considerably less leeway for manual tuning, as its voltage may be set to a maximum of 1.025 volts (XT: 1.15 volts) and the GPU power limit to a maximum of 230 watts (XT: 293 watts). For this reason, we asked Nvidia for the following opinion on overclocking from Founders Editions: Anything you can set in tuning tools such as the Nvidia Inspector, MSI Afterburner, EVGA Precision & Co. is not a warranty violation. This is only applicable when you modify the firmware of your graphics card. In this sense: fire free!

Radeon RX 6800 (XT) im Test: Kühlung, Lautheit, Verbrauch, Effizienz (pcgameshardware.de)

 

[Jawed]

How long will Ampere FE be on sale?

 

[DegustatoR]

How long will Ampere FE be on sale?

Until the next generation of GPUs, as it was with all previous FEs from NV.

 

[Kaotik]

Until the next generation of GPUs, as it was with all previous FEs from NV.

Limited sales at couple US retailers isn't really being for sale. RTX 3070 FE has never been available in majority of the world.

 

[xpea]

Limited sales at couple US retailers isn't really being for sale. RTX 3070 FE has never been available in majority of the world.

Well FE may have been limited but not the AIB cards. I live in China and plenty of different RTX 3070 are easily available on Taobao and JD. Same as Honk Kong and Vietnam where my friends got theirs. Most of Asian countries have no issue with 3070. In fact, the supply has been solid on this model.

 

[pcchen]

Well FE may have been limited but not the AIB cards. I live in China and plenty of different RTX 3070 are easily available on Taobao and JD. Same as Honk Kong and Vietnam where my friends got theirs. Most of Asian countries have no issue with 3070. In fact, the supply has been solid on this model.

Yeah, it's similar here in Taiwan. A large computer retailer here has 3070 and 3090 mostly available, only 3080 and 6800 are in very tight supply. 6800XT is completely out of stock, though. Price-wise, it's a bit more expensive but not by very much (e.g. ASUS DUAL fan 3070 @ NT$16,990 ~ US$600 including tax)

 

[Deleted member 2197]

GeForce RTX 3080 20GB Registered at EEC - Coming in December?

Yes, no, yes, no ... that's pretty much what the rumor train has been on an RTX 3080 fitted with 20GB VRAM. Face it, there has been chatter on a 20GB model for ages now, but things get serious once they get registered, and that happened once again on the EEC.

Looking at the dates, it cannot be coincidental; a decision was made shortly after AMD released its Radeon RX 6800, prompting NVIDIA to abandon its plans for the 20GB model and also for the 16GB RTX 3070. It's an MSI entry unveiled a 20GB variant for the RTX 3080 at the Eurasian Economic Commission (EEC).

GeForce RTX 3080 20GB Registered at EEC - Coming in December? (guru3d.com)

 

[Xmas]

This "ML all the stuff!" approach though just doesn't fundamentally make any sense. If you need entirely known, perfectly predictable results you wouldn't use an essentially statistics based approach to begin with. It's the same reason "neural rendering" is nigh abandoned at this point already. You don't need to guess the rendering equation, you have it and need to do it as fast as possible.

It's worth pointing out that for many rendering tasks the rendering equation is not the goal, just a step on the road of creating a certain impression in the human viewer. That makes them quite different from simulation tasks.