Nvidia Claims Higher RTX Frame Rates With DLSS Using 3DMark Port Royal Benchmarks

Giroro

Splendid


Never... because by the time Devs figure out how to use it Nvidia will have mothballed DLSS for whatever next year's anti-aliasing gimmick is. Just like every other "new/better/faster" proprietary AA acronym Nvidia has marketed with every new graphics card line in recent memory.
 

AlistairAB

Distinguished
May 21, 2014
229
60
18,760
A machine learning approach to a non-interactive benchmark... so "1440p" DLSS is nVidia lying, most likely based on 1080p native, and you don't control the camera making it ridiculously simple. Not impressed as usual with DLSS.

I really wish nVidia would stop labeling their AA method based on the target upscale, and base it on the native resolution like everyone else.
 
Feb 4, 2019
1
0
10
Any day now I'll be able to enjoy the vast improvement in VR performance thanks to the Simultaneous Multi-Projection on my 1080ti. Right? Soon?
 

AlistairAB

Distinguished
May 21, 2014
229
60
18,760
This actually shows a useful niche for DLSS: In game cut-scenes. Since the camera is fixed, the training is pretty effective. This looks much better than the Infiltrator demo, where I preferred the TAA. I still think DLSS will most likely be useless for gaming, but for a cut-scene, that would work. With DLSS you are basically downloading extra info, so it takes more storage than complete real time, but way less than full recorded video. You could run an in game cut-scene at a lower resolution but have it impress with visuals and clarity. I'm trying to be positive here :)

I do want to point out that it is still a wasteful technology. The DLSS technology required die space on the GPU and extra cost in the data center. Much better would have been to invest that money in enlarging the game performance. The RTX 2060 with the large die could equal the 2080 ti in game performance if the 2080 ti was solely devoted to traditional raster performance. The die size ballooned and is inefficient with RTX.
 

AlistairAB

Distinguished
May 21, 2014
229
60
18,760
RTX 2060 is 445mm2 at 12nm. 1080 Ti is 471mm2 at 16nm. We know that from the Radeon 7, we can expect at least a 40 percent reduction in die size from 7nm. So the 1080 ti at 7nm would have been about 471mm2 * 0.6 = 282 mm2.

2080 Ti performance (clock speed increases of at least 20 percent) at 7nm with only 282 mm2. Without the wasteful RT and DLSS technology added to the die...

Vega 7 is 331mm2, so the 2080 ti equivalent 1080 ti at 7nm would be much cheaper to make than Vega 7, and much faster. We all know nVidia is trying to push their own tech instead of giving us what we want...
 
Wonder if anybody could actually work out a reasonably accurate cost / benefit ratio for adding the Tensor cores to the GPU die, VS just using that space for more general purpose cores?

The problem with adding the custom cores to the GPU is, all of your development work for those cores gets thrown away as soon as they get left out of the GPU, unless of course your code can also run on the general compute cores, which then brings the custom cores benefit into question.

So, how long is NVIDIA going to require consumers to pay for cores they may never get full use of? Replacing those Tensor cores with the general purpose cores would work for every single game. Will we see Tensor cores on all new GPUs from NVIDIA going forward, or are they segmenting their own market?
 
Alistairab,
First off, who cares if it's 1080p upscaled to 1440p if the end result is that it LOOKS like 1440p but runs at a higher FPS?

If NVidia put (1080p + DLSS) vs (1440p) then people would then scream "but, but, but it's not the same resolution" so I'm not sure what exactly NVidia's supposed to do about that.

DLSS isn't going anywhere likely. Ground truth analysis is pretty amazing.

As for the DIE SPACE being better used for traditional rasterization? You don't think NVidia's engineers are very aware of this? The PROBLEM is that rasterization has its limits and at some point the hardware needs to change so you can introduce some Ray-Tracing etc that can do things rasterization just can't do.

AI-based optimizations, ray-tracing additions, and other methods that can make better use of the added processing elements are the future. We should get that to varying degrees with the next-gen consoles using AMD's NAVI too and I expect NAVI on desktop to be similar to NVidia's RTX but probably less proportional die space for allocated to the newer compute units.

GAMING won't benefit much from DLSS?
There's no evidence to support that if you understand how the process works.

Finally, don't forget that the newer compute units (INT32, Tensor, RT) are also part of CUDA v10 so when some applications like VIDEO EDITING programs optimize for it there's the potential for a huge boost in performance compared to the same die size with traditional shader cores (i.e. NVidia Pascal)... I'm not certain but I think Turing GTX has INT32 but not RT and Tensor.
 

AlistairAB

Distinguished
May 21, 2014
229
60
18,760


I agree with a lot of what you said, but I don't think anyone would be upset if NVidia called it 1080p DLSS, since that is what it is. NVidia calling it 1440p is intentional to obscure the computational cost of DLSS. The point is DLSS was invented to bring clarity back to blurry ray tracing, since ray tracing won't work clearly and at high resolution for a long time.

Problem is: I don't have enough performance to play Anthem at 4k for example, cheaply. We are far from needing ray tracing performance, but we still desperately need more and cheap rasterization performance. I'd be very interested in an nVidia ray tracing console however, if new games were made for it exclusively. The Quake demo was interesting.

 

jamesltoa4

Distinguished
Feb 20, 2009
16
0
18,510
AFAIK most reasonably recent systems such as those that would allow a PCIe connection, and run an OS such as Windows 10 do not have a "Boot Drive"
They have (as can be seen using the Storage ~ Disk Management links from the "This PC" icon
PARTITIONS labelled "System" "Boot"
The OS being run is only associated with the "Boot" facility by the OS selection list and startup modules.
That being why you can have GRUB and WIN-10 partitions on the same drive in a system.

So – where the Boot is, and it's size is only relevant in the short period until the OS startup process actually starts running.

Regarding the terminology
There is no such thing as a C: drive
Almost all storage facilities – hard drive or memory stick or SSD are IDE devices –
As in the drive comes with "Integrated Drive Electronics"
So PATA and SATA and PCI-E devices are all IDE

Additionally, there is NO requirement for the windows OS to actually run from a PARTITION labelled C:
When the startup process runs it can allocate any letter from Z down to C to any PARTITION including the one that the selected OS is running from.

You purport to have some 'Technical' expertise, so please show it by using the appropriate technology.

That dealt with – regarding actual devices in a system
Windows 10 will run on a 32GB partition –
You will of course have problems running any modern apps, or even Windows Update with just 32GB
I believe you should, for the current windows 10 management and change it for you Microsoft policy, need to have at least 64GB for the OS itself
With, maybe an additional 30+GB needed when you get to the bi-yearly update process.
So – anyone buying a system will need a 120GB level PARTITION for the OS
Regarding "Boot" – well maybe 200MB

Now, if you want to run applications or games in a fast response mode – then separate their storage from the OS Pagefile
As in a separate interface path to the motherboard
That should allow loading of programs from the drive concurrently with the OS using the Pagefile.
OK better to have enough RAM to allow you to not to have any Pagefile.

Re OS partition separate from the apps and data –
Yes – definitely not only for backup and restore purposes but also for throughput speed
allow OS processes to run on 1 device with the app (especially games) running from a separate device
Consider file processing –
OS gets asked for a file – is in in RAM cache? Is it in Pagefile cache?
Where on the device is it – scan MFT – repeat for the MFT pages – in RAM, on Pagefile)
Allocate RAM for the data block if not already done – as in page out somethings to get re-assignable RAM
Know where the file is on storage – Logical Block Nº request blocks from the device –
Is that block in cache on the device – if not load it into cache – and pass it back to the OS.

Yes - Real RAM can so much improve performance as can (if you have one) Pagefile on a fact response device.

So –
Do you need a fast BOOT PARTITION – Nope – BOOT is only done before system startup
Should you have your OS running from a fast partition – Yes it handles the I/O
Should you have your Pagefile on a fast partition – Yes if you cannot avoid one - it delays almost all throughput
Should you have the app and data separate from the OS and Pagefile – Yes especially if the device has a large cache – as in the up to 32GB optane
I would (am) looking at a system for good response under Windows 10 as
RAM 16GB at least – preferably for VM – as in sandbox and VM browsing may need 16GB for a session
CPU – at least 4 cores – see the current apps from MS
OS partition – SSD is nice – but if 32GB of Optane is available – then won't that hold almost all the files used by the OS so reducing access to the drive storage itself.
Apps partition – again Optane – another 32GB would be nice – but games etc may well have over 32GB of program and essential data – screen environment etc.
Multimedia and backup – only written once and maybe read once – and sped – well, at what multiplication rate can you watch and listen.

So – for a fast system 4+ core, 32GB RAM, 32GB optane on OS device 32GB Optane on apps device and large bulk store for installed apps and other bulk data

Boot drive – who cares as long as it does not also hold the OS PARTITION.
 

mjbn1977

Distinguished
Well, I ran the Port Royal benchmark with DLSS yesterday. My RTX 2080 were able to get 40fps vs 28fps without DLSS. One thing that I noticed was that with DLSS on the graphics looked much more crisp and sharp than without it. If this performance increase in combination of the better quality picture will be like this in real world gaming, then this feature actually could be a hit.
 

mjbn1977

Distinguished
By the way....DLSS is not conventional up-scaling. Even so the picture is rendered at a a lower resolution you will still get a real higher resolution picture in the end. It is much higher quality than normal up-scaled pictures. The AI is basically filling in the gaps and creates a real non-up-scaled picture. I think we have to think different when talking up-scaling and DLSS. Normal up-scaling is a bad thing with mediocre results. DLSS up-scaling is a completely different approach with pretty incredible results, judging the examples I have seen so far. If the pictures are as crisp as in the demos and Port Royal, someone would almost be stupid not to use DLSS.
 

I haven't looked over any Port Royale DLSS comparisons yet, but based on what I saw in the prior FFXV demo comparisons, it doesn't actually seem to be "better image quality". The DLSS process appears to effectively apply a sharpening filter over the entire scene. This might make the image appear sharper, but a lot of what's getting sharpened is actually intended to be out of focus, making this over-aggressive sharpening a rendering artifact more than anything. Things like depth of field and atmospheric effects are getting filtered out of the scene, artificially making things look "sharper" that are not supposed to be. Meanwhile, parts of the scene that were intended to be sharp and in focus are actually still blurrier than they would have been if rendered at native resolution without aggressive AA. Perhaps the quality of implementation might vary on a per-game basis, but that previous demo left me unimpressed.

I suppose DLSS could be considered okay as another way to upscale lower resolutions, which could make hybrid raytracing somewhat more usable, but one needs to take a close look at what it's actually doing to image quality, and I think that's something that might often get ignored in performance comparisons.
 

Prizm4

Honorable
Oct 28, 2015
6
1
10,510
We’re missing some information here. Those graphs are nonsense. The RTX 2060 has a performance gain of 50% compared to what? No AA at all?? And how does that second graph relate? For the 2060, It shows 28fps with DLSS on and 19fps with DLSS off. That’s not a 50% difference ??
 

mjbn1977

Distinguished


Ok, ok...it's more like 47.5%.....but close enough to call it 50% performance gain.

Math refresher:
19/100=0.19
0.19*147.5=28.025
or simpler:
19*1.475=28.025
or to be exact:
28-19=9
9/19*100=47.37%

Oh, by the way. It is comparing it to the same benchmark (all settings the same) but with TAA instead of DLSS.

 
Wow... It looks like DLSS actually offers both worse image quality and worse performance than what can be had by simply upscaling from a lower resolution with TAA in actual games. I can see why Nvidia didn't want this to be in games until after reviews were out for all of the RTX cards. It seems as though this might just be a useless feature, trying to find some purpose for those Tensor cores left over from their compute cards, but failing to accomplish anything better than what can be done without them...

View: https://www.youtube.com/watch?v=3DOGA2_GETQ
 

hannibal

Distinguished
Yeah... DLSS works in still pictures where image orientation does not change or when the camera moves exacly same route every time. So it is good in cut scenes and maybe in game bensmark tests and normal bensmark tests. But when used in games where scene is different each time you play... the image is worse than in normal upscaling...
really usefull if you run in your Computer only benchmarks and demos... not real games.