Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion.

Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules.

 

Go Back   Doom9's Forum > Hardware & Software > Software players

Reply
 
Thread Tools Search this Thread Display Modes
Old 2nd September 2020, 06:24   #60041  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
with current information it's a yes.

nvidia talks about 1.9 times(context as "always" with nvidia is not really given) it looks more like.

a 220 watt GPU is faster then a 250 watt so 1.2-1.5.
huhn is offline   Reply With Quote
Old 2nd September 2020, 10:00   #60042  |  Link
chros
Registered User
 
chros's Avatar
 
Join Date: Mar 2002
Posts: 2,323
Quote:
Originally Posted by huhn View Post
beware our current knowable about the 3000 series and it's FP32 performance is based on the fact that one core can do 2 FP32 operation instead of 1 like turing. proper testing is need to see if this works as advertised and doesn't behave like "hyperthreading" or more like bullldozer.
Ohhh, that's why the cuda cores in the leak the other day are exactly the half that nvidia states! Thank for this!
Am I the only one who feel bullshit in the air? It's like amd would say that a 6 cores SMT CPU is 12 cores. And if that's the case then why nvidia felt the need to lie like this, instead of put out good marketing for "hyperthreading"?
I guess we will have to wait for benchmarks...
__________________
Ryzen 5 2600,Asus Prime b450-Plus,16GB,MSI GTX 1060 Gaming X 6GB(v398.18),Win10 LTSC 1809,MPC-BEx64+LAV+MadVR,Yamaha RX-A870,LG OLED77G2(2160p@23/24/25/29/30/50/59/60Hz) | madvr config

Last edited by chros; 2nd September 2020 at 10:06.
chros is offline   Reply With Quote
Old 2nd September 2020, 10:16   #60043  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
the more you look at it the clearer it get's it 3090 has 2 set of ~5k cuda cores in turing one set was INT32 and the other normal CUDA FP32 in ampare they seem to be able to switch between INT32 and FP32 i have no clue if all core can do that or only half.

we will know when official data is available with the INT32 performance.
huhn is offline   Reply With Quote
Old 2nd September 2020, 12:29   #60044  |  Link
nevcairiel
Registered Developer
 
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,342
The "CUDA cores" were always a bit of theoretical thing, its not like you can access them individually. But if it can handle twice as many operations per tick, does that really conflict with saying it has twice as many cores?
__________________
LAV Filters - open source ffmpeg based media splitter and decoders
nevcairiel is offline   Reply With Quote
Old 2nd September 2020, 13:10   #60045  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
if you get the information of FP32x2 and 10K cores do you now have to expect the performance of 20k turing cuda cores?

beware these are not on the same spec sheet!
huhn is offline   Reply With Quote
Old 2nd September 2020, 14:35   #60046  |  Link
SirMaster
Registered User
 
Join Date: Feb 2019
Posts: 231
The CUDA core definition has not changed from Turing to Ampere.

But the scenarios where you can utilize 100% of CUDA cores are sometimes limited as the streaming multiprocessor has been reworked to do FP+FP, FP+INT, etc., rather than FP *or* INT.
SirMaster is offline   Reply With Quote
Old 2nd September 2020, 15:00   #60047  |  Link
NoTechi
Registered User
 
Join Date: Mar 2018
Location: Germany
Posts: 93
Quote:
Originally Posted by huhn View Post
madshi "fixed" 3D for new driver many month ago he simply never released the build.
This is interesting because I assumed that it would not work without nvidia 3D driver support. Then I look forward to a new madvr 3D with nvidia re-enabled version

NoTechi
NoTechi is offline   Reply With Quote
Old 2nd September 2020, 15:36   #60048  |  Link
chros
Registered User
 
chros's Avatar
 
Join Date: Mar 2002
Posts: 2,323
Quote:
Originally Posted by nevcairiel View Post
The "CUDA cores" were always a bit of theoretical thing, its not like you can access them individually. But if it can handle twice as many operations per tick, does that really conflict with saying it has twice as many cores?
I think it does:
Quote:
Originally Posted by chros View Post
It's like amd would say that a 6 cores SMT CPU is 12 cores.
__________________
Ryzen 5 2600,Asus Prime b450-Plus,16GB,MSI GTX 1060 Gaming X 6GB(v398.18),Win10 LTSC 1809,MPC-BEx64+LAV+MadVR,Yamaha RX-A870,LG OLED77G2(2160p@23/24/25/29/30/50/59/60Hz) | madvr config
chros is offline   Reply With Quote
Old 2nd September 2020, 16:01   #60049  |  Link
nevcairiel
Registered Developer
 
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,342
Quote:
Originally Posted by chros View Post
I think it does:
No, thats hardly the same thing, because SMT doesn't actually mean it can run twice the instructions. Its an apples to oranges comparison entirely.
The CUDA core layout was reportedly changed so that it actually has hardware to run two FP32 computations, instead of just one.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders
nevcairiel is offline   Reply With Quote
Old 2nd September 2020, 17:26   #60050  |  Link
chros
Registered User
 
chros's Avatar
 
Join Date: Mar 2002
Posts: 2,323
Hm, cheers.
__________________
Ryzen 5 2600,Asus Prime b450-Plus,16GB,MSI GTX 1060 Gaming X 6GB(v398.18),Win10 LTSC 1809,MPC-BEx64+LAV+MadVR,Yamaha RX-A870,LG OLED77G2(2160p@23/24/25/29/30/50/59/60Hz) | madvr config
chros is offline   Reply With Quote
Old 2nd September 2020, 19:38   #60051  |  Link
*MidnightWatcher*
Registered User
 
*MidnightWatcher*'s Avatar
 
Join Date: Jun 2020
Posts: 67
The question for me now is this -- Should I pick up a cheap used RTX 2080 or spend more and buy an RTX 3070? What settings would an RTX 2080 not be able to do that the 3070 would?
*MidnightWatcher* is offline   Reply With Quote
Old 2nd September 2020, 20:00   #60052  |  Link
Alexkral
Registered User
 
Join Date: Oct 2018
Posts: 319
Quote:
Originally Posted by *MidnightWatcher* View Post
What settings would an RTX 2080 not be able to do that the 3070 would?
The first thing I always think of when I see these kinds of questions is frame interpolation. In addition to be partly done on the GPU, it is not the same to process with madVR at 24 fps than at 60 or even more.
Alexkral is offline   Reply With Quote
Old 2nd September 2020, 20:32   #60053  |  Link
chros
Registered User
 
chros's Avatar
 
Join Date: Mar 2002
Posts: 2,323
Quote:
Originally Posted by *MidnightWatcher* View Post
The question for me now is this -- Should I pick up a cheap used RTX 2080 or spend more and buy an RTX 3070? What settings would an RTX 2080 not be able to do that the 3070 would?
Which GPU do you have now?
What do you need it for? Gaming? then yes. Do you need the new inventions, HDMI 2.1/AV1? yes, again.
Would you only use it for madvr? then good question. Do you need RTX for it at all?! A used Pascal 1080/Ti would be cheaper
There's no point "future proofing" unless you need those features right now.

Quote:
Originally Posted by Alexkral View Post
The first thing I always think of when I see these kinds of questions is frame interpolation. In addition to be partly done on the GPU, it is not the same to process with madVR at 24 fps than at 60 or even more.
Yes, but personally I don't need this.
__________________
Ryzen 5 2600,Asus Prime b450-Plus,16GB,MSI GTX 1060 Gaming X 6GB(v398.18),Win10 LTSC 1809,MPC-BEx64+LAV+MadVR,Yamaha RX-A870,LG OLED77G2(2160p@23/24/25/29/30/50/59/60Hz) | madvr config
chros is offline   Reply With Quote
Old 2nd September 2020, 20:46   #60054  |  Link
SamuriHL
Registered User
 
SamuriHL's Avatar
 
Join Date: May 2004
Posts: 5,351
If you're upscaling lower resolutions at 60fps then faster speeds are warranted. How much so depends on far too many factors for anyone to realistically answer. I would also not buy an RTX card banking on the fact that madshi is going to be building new tensor core algorithms. He absolutely IS building new tensor core algorithms. However, we have no guarantees at all that they will ever see the light of day in madvr on a PC. They may, but, that's not a selling point of a new card. So right now, all that matters is raw speed. I personally use copyback D3D because I want menus when I stick a disc in the machine. The faster memory speeds on the new 30xx series may or may not improve things on that front. It remains to be seen if the 2xfp32 will have a real impact on what can be done in madvr, as well. These are all unknowns until we get the cards in our hands. But ultimately what card is right for you depends on your needs/wants/desires and what content you will be viewing and on what type of screen. I'm running a 2070 right now which does 99% of what I want. If I wasn't also into gaming and looking to avoid picking up a new console, I might not even bother chasing after the use case I'm looking at solving with a new card and just stick with the 2070. But I weighed the options for myself, and I see the 30xx cards as being a very useful upgrade overall. YMMV
__________________
HTPC: Windows 11, AMD 5900X, RTX 3080, Pioneer Elite VSX-LX303, LG G2 77" OLED
SamuriHL is offline   Reply With Quote
Old 2nd September 2020, 23:00   #60055  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
Quote:
Originally Posted by nevcairiel View Post
No, thats hardly the same thing, because SMT doesn't actually mean it can run twice the instructions. Its an apples to oranges comparison entirely.
The CUDA core layout was reportedly changed so that it actually has hardware to run two FP32 computations, instead of just one.
then doubling the number of available cores and saying a "core" can do FP32x2 is easily misleading. because it really looks like it can do FP32+FP32 or FP32+INT32 where turing was always able to do FP32+INT32 (INT 32 where independent cores) this is a great update if true.

they have double the CUDA core per SU or they can do FP32x2 i have a hard time believing they can do both at the same time.
huhn is offline   Reply With Quote
Old 2nd September 2020, 23:53   #60056  |  Link
*MidnightWatcher*
Registered User
 
*MidnightWatcher*'s Avatar
 
Join Date: Jun 2020
Posts: 67
Quote:
Originally Posted by chros View Post
Which GPU do you have now?
What do you need it for? Gaming? then yes. Do you need the new inventions, HDMI 2.1/AV1? yes, again.
Would you only use it for madvr? then good question. Do you need RTX for it at all?! A used Pascal 1080/Ti would be cheaper
There's no point "future proofing" unless you need those features right now.
No gaming, using madVR only. My current GPU is a 1060 6GB and it does about 75% of what I want. If I upgrade I'd go for an RTX model for the tensor cores, just in case madshi ever (fingers crossed) released an update that took advantage of it somehow.
*MidnightWatcher* is offline   Reply With Quote
Old 2nd September 2020, 23:59   #60057  |  Link
Stereodude
Registered User
 
Join Date: Dec 2002
Location: Region 0
Posts: 1,436
I have a 1080 non-Ti and I can barely play back 2160p60 to a UHD TV without dropping frames using madVR. I had to dial things way back to like bicubic chroma upscaling and most other stuff turned off. I wouldn't consider a Pascal card if that's what you want to do.

Edit: I should add that I couldn't find settings to let me play back HDR 2160p60 without dropped frames if it doesn't have black bars. SDR is okay.

Last edited by Stereodude; 3rd September 2020 at 00:07.
Stereodude is offline   Reply With Quote
Old 3rd September 2020, 00:25   #60058  |  Link
jkauff
Registered User
 
Join Date: Oct 2012
Location: Akron, OH
Posts: 491
Quote:
Originally Posted by Stereodude View Post
I have a 1080 non-Ti and I can barely play back 2160p60 to a UHD TV without dropping frames using madVR. I had to dial things way back to like bicubic chroma upscaling and most other stuff turned off. I wouldn't consider a Pascal card if that's what you want to do.
Hmm. I have a 1080 in my HTPC attached to my 4K HDR television, running madVR with custom optimized resolutions. I can easily upscale my Blu-ray movies with NGU High.

Is there any reason you're using 2160p60? Doesn't your TV support lower resolutions?
jkauff is offline   Reply With Quote
Old 3rd September 2020, 00:42   #60059  |  Link
VBB
Registered User
 
VBB's Avatar
 
Join Date: May 2016
Location: Long Beach, CA, USA
Posts: 620
He must be talking about 60fps content, like Gemini Man, and also using tone mapping at the same time. I don't use tone mapping, but my GTX 960 can easily play such content.
__________________
Henry | LG OLED65C7P | Denon AVR-X3500H | ELAC Uni-Fi x7 | ELAC Debut 2.0 SUB3030 x2 | NVIDIA SHIELD TV Pro 2019 | Plex
VBB is offline   Reply With Quote
Old 3rd September 2020, 00:46   #60060  |  Link
Asmodian
Registered User
 
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
Quote:
Originally Posted by huhn View Post
they have double the CUDA core per SU or they can do FP32x2 i have a hard time believing they can do both at the same time.
I agree, I think the number of CUDA cores reported is actually the number of FP32 units, not cores that can each do two FP32 operations. So these new "cores" are weaker than the old cores because they cannot do INT32 at the same time, more of an expansion of the old INT32 units into FP32 units than really doubling everything in what they used to call a CUDA core. Of course, marketing is trying to make it sound good and in the process of explaining what they did technically they end up saying two things that are mutually exclusive.

In this case I would probably do the same thing. If you double the FP32 units you have to double the reported core count, everyone knows each core is a FP32 unit. But then architecturally the way you got double the FP32 units did not double everything traditionally included in a CUDA core. With no good way to define GPU "cores" in the first place, and consumers not getting excited by architectural details, Nvidia had no option but to be confusing... and they picked the confusion that made them look better.

But this is not like SMT either, there really is twice the hardware we traditionally cared about in a GPU core. It is not a doubling of the front end to better utilize idle hardware or similar.
__________________
madVR options explained
Asmodian is offline   Reply With Quote
Reply

Tags
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not 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

Forum Jump


All times are GMT +1. The time now is 15:25.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.