[Techpowerup] NVIDIA DLSS and its Surprising Resolution Limitations - Page 9 - Overclock.net - An Overclocking Community

Forum Jump: 

[Techpowerup] NVIDIA DLSS and its Surprising Resolution Limitations

Reply
 
Thread Tools
post #81 of 91 (permalink) Old 02-20-2019, 09:34 AM
Looking Ahead
 
TheBlademaster01's Avatar
 
Join Date: Dec 2008
Location: Cluain Dolcáin, Leinster (Ireland)
Posts: 13,043
Rep: 785 (Unique: 536)
Good point, coincidentally he also missed a couple of SMs on GP100 (so that might cancel both errors out a bit). HBM2 doesn't matter since the derivation of the RT/Tensor Cores in terms of CUDA and TMUs is based on two HBM2 GPUs.

The main issues IMO are with the differences in SM composition (TU102 lacking FP64 units and not knowing the ratio between Tensor and RT cores). And indeed how well the 12nm node scales relative to 16nm.

 



TheBlademaster01 is offline  
Sponsored Links
Advertisement
 
post #82 of 91 (permalink) Old 02-20-2019, 09:40 AM
Waiting for 7nm EUV
 
tpi2007's Avatar
 
Join Date: Nov 2010
Posts: 11,382
Rep: 894 (Unique: 503)
There is a site that from time to time posts detailed views of dies using speciality equipment (electron microscopes, if I'm not mistaken). I don't recall the name right now. That could eventually help if they happened to get their hands on Volta and Turing.


tpi2007 is offline  
post #83 of 91 (permalink) Old 02-20-2019, 09:46 AM
New to Overclock.net
 
Leopardi's Avatar
 
Join Date: Dec 2007
Posts: 1,559
Rep: 18 (Unique: 18)
Quote: Originally Posted by ILoveHighDPI View Post
https://www.techpowerup.com/252550/n...on-limitations



This is the final nail in the coffin.
My primary complaint with DLSS is that the existence of the Tenser Cores on a chip where you are not going to be using Ray Tracing is fundamentally wasteful, that Silicon Area should just be used for more CUDA cores and Nvidia needs to have a high end GTX card available.
Secondly we have seen that basic up-scaling gives almost exactly the same benefits as DLSS, but does so without any need for extra hardware: https://www.techspot.com/article/1712-nvidia-dlss/

Still the idea of DLSS has been lucrative, a 50% boost to framerate with only a small cost to image quality. Many people would have gladly taken that compromise to reach higher framerates...

Only today we find out that DLSS is itself a bottleneck to reaching higher framerates.
The 25% scaling + 2xMSAA technique that was used in Rainbow Six: Siege was equal to having a bit lower resolution on a CRT. Much better than this DLSS will ever be, the image was razor native sharp without any loss or ghosting artifacts.

Leopardi is offline  
Sponsored Links
Advertisement
 
post #84 of 91 (permalink) Old 02-20-2019, 09:56 AM
New to Overclock.net
 
doritos93's Avatar
 
Join Date: Oct 2009
Location: Montreal
Posts: 2,309
Rep: 119 (Unique: 84)
So HWU has proven that this is garbage tech ATM. Can we stop calling this a feature now



doritos93 is offline  
post #85 of 91 (permalink) Old 02-20-2019, 10:59 AM
Waiting for 7nm EUV
 
tpi2007's Avatar
 
Join Date: Nov 2010
Posts: 11,382
Rep: 894 (Unique: 503)
Quote: Originally Posted by doritos93 View Post
So HWU has proven that this is garbage tech ATM. Can we stop calling this a feature now
https://youtu.be/3DOGA2_GETQ

I made a thread about it, including that video and the accompanying written article:

https://www.overclock.net/forum/226-...delivered.html


tpi2007 is offline  
post #86 of 91 (permalink) Old 02-20-2019, 01:55 PM
New to Overclock.net
 
DNMock's Avatar
 
Join Date: Jul 2014
Location: Dallas
Posts: 3,482
Rep: 171 (Unique: 125)
Quote: Originally Posted by tpi2007 View Post
The 12nm that they are using is just a slight improvement over 16nm, it probably doesn't amount to that much. Also, GV100 actually has 84 SMs, but we only ever saw products with 80 enabled due to yield and probably also power consumption reasons, so take that into account in your math. Also, there is a difference in that GV100 uses an HBM2 memory controller, thus saving a bit of die space compared to the GDDR6 memory controller equipped Turing dies.
Oh I'm sure there are a heck of a lot of things I got wrong there. Wouldn't be surprised if my basic math was wrong in spots lol.


DNMock is offline  
post #87 of 91 (permalink) Old 02-20-2019, 05:33 PM
New to Overclock.net
 
Join Date: Apr 2018
Posts: 124
Rep: 0
Quote: Originally Posted by tpi2007 View Post
Maybe they quietly changed their minds, they are masters at that. Or maybe they intend to enforce it later on, but eased the system for these first few titles, given the general backlash against Turing's high prices and lackluster RTX performance, to avoid having people complain about one more walled garden barrier to entry to use the new features, even more so after having waited for so long to see them in action in actual games.

As I said back then when the sites quoted above reported on it, there is no technical reason for Nvidia to require GFE to get the game specific DLSS packages, it's a telemetry and marketing move for them above all. If the DLSS packages are small they can be packed with the game ready drivers like SLI profiles; if they are relatively big, like 100 MB of more (seems more likely), they can be optional downloads from their site next to the drivers or, much more practical and with a lot more sense to it, they can ship the finished package to the game dev and then it gets shipped with the game at launch or as part of an update - people already download several GB's worth of game updates, a DLSS package that is 100-200 MB big is a drop in the ocean.

Pretty sure the DLSS profile is part of the driver, otherwise those of us without GFE wouldn't be able to use DLSS. Nvidia is probably just pushing GFE as a automatic driver updater again.
Crinn is offline  
post #88 of 91 (permalink) Old 02-20-2019, 06:05 PM
Waiting for 7nm EUV
 
tpi2007's Avatar
 
Join Date: Nov 2010
Posts: 11,382
Rep: 894 (Unique: 503)
Quote: Originally Posted by Crinn View Post
Pretty sure the DLSS profile is part of the driver, otherwise those of us without GFE wouldn't be able to use DLSS. Nvidia is probably just pushing GFE as a automatic driver updater again.

Right now it's part of either the driver or the games themselves, but beware that they didn't erase that paragraph I quoted from the Turing whitepaper, so once there is more than a handful of games (right now we have three games and one benchmark), they might try to bring it under the GFE umbrella by claiming that they can use the same Deep Neural Network model for more than one game, thus they would phrase this centralization as an efficiency measure to justify the move. It's a possible justification if you read the whitepaper and incidentally what some claimed back in August of last year to justify GFE handling DLSS profiles.

That's all assuming that they can turn the DLSS trainwreck around first and make it something worthwhile, otherwise people won't care whether or where it's available.



Last edited by tpi2007; 02-20-2019 at 06:11 PM.
tpi2007 is offline  
post #89 of 91 (permalink) Old 02-20-2019, 07:01 PM
New to Overclock.net
 
Kaltenbrunner's Avatar
 
Join Date: Dec 2011
Posts: 2,272
Rep: 17 (Unique: 16)
Quote: Originally Posted by Telimektar View Post
What can you do with those RTX and Tensor cores besides DLSS and RTX effects, I'm still not clear on that ? I wonder if console and arcade emulators could use them in some way.
In typical games I guess not much. Nviada used them so that they can make a GPU they could sell to gamer's AND to scientists/etc

What is DLSS? Is it new AA ?

Kaltenbrunner is offline  
post #90 of 91 (permalink) Old 02-20-2019, 07:21 PM
Smug, Jaded, Enervated.
 
GHADthc's Avatar
 
Join Date: Aug 2012
Location: Terra Australis
Posts: 1,269
Rep: 101 (Unique: 66)
Quote: Originally Posted by ToTheSun! View Post
The AI developed a conscience and started caring about the environment and energy expenditure. It also reduces framerate because it heard Jensen saying that amount of performance is irresponsible.
Man...if we could still REP posts...
GHADthc is offline  
Reply

Quick Reply
Message:
Options

Register Now

In order to be able to post messages on the Overclock.net - An Overclocking Community forums, you must first register.
Please enter your desired user name, your email address and other required details in the form below.
User Name:
If you do not want to register, fill this field only and the name will be used as user name for your post.
Password
Please enter a password for your user account. Note that passwords are case-sensitive.
Password:
Confirm Password:
Email Address
Please enter a valid email address for yourself.
Email Address:

Log-in



Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)
 
Thread Tools
Show Printable Version Show Printable Version
Email this Page Email this Page


Forum Jump: 

Posting Rules  
You may post new threads
You may post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are Off
Pingbacks are Off
Refbacks are Off