2080 VS 1070 SLI benchmarks. A true Side-grade

[reptileexperts]

So I wanted to get away from SLI to cool down my system a bit. Maybe prepare for a water loop again and clean things up. Bring on the new RTX cards, no benchmarks aside, I pre-ordered the 2080ti knowing it would still be the highest end card on the market. . . fast forward to benchmarks and the dollar to frame ratio was simply insane, so I went ahead and returned my 2080ti and picked up a reference 2080 for a fraction of the cost thanks to a birthday discount at best buy. My current rig was a SLI 1070 Strix edition setup, paired with a water cooled AIO 8700k @ 5ghz at 1.4V. I have had no issues with SLI scaling by in large, but again wanted to cool things down and try a single card solution. 

 

So what did I find? It's a side grade. I knew it would be since 1070 SLI goes head to head and beats out 1080ti's on certain tests. The 2080 performed similar to a 1080ti so this should be the same. 

 

In shadows of the tomb raider benchmark, fully maxed out at 1440p, the 2080 eeks past the 1070 SLI despite near perfect SLI scaling

Makes me happy, since I know this is what most newer games will compare to. 

 

Going to synthetic benches, we see a slightly different story.

 

Oddly enough, the new drivers reduced my typical ultra score by around 500 points as my system would typically be above 8k. .  . but this was tit for tat with current drivers at current system settings. SLI rig destroys the 2080

Time Spy, again 1070 SLI comes out dominating the competition. The score in the background was the non-official benchmark pre-release. My score became much closer. 

More realistic to what we saw in tomb raider, but still showing synthetic benchmarks giving perfect SLI scaling. 

 

All in all I am happy with my results, and will go ahead and SLI NVI-Link 2080's in the very near future once i sell my 1070's to cover the second card. The single card performance is beautiful, and the system is extremely quiet under load. The fans never get past 2100 RPM, which keeps things quiet and cool-ish holding the temps around 81 C during intensive gaming. Seems to be the common temperature among all the reviews that's been posted. 

 

As far as overclock goes. . . 

 

I was successful after some tinkering of hitting +170 on Core, +320 on memory. This was a manual overclock as scanner has not yet been released? (that I've found)

 

Love or hate RTX. even the 2080 is an insane card. 

 

 

[EOZ]

GTX 1080Ti is better deal than RTX 2080

[Firewrath9]

nice.

im sad and stuck with a single 1070. I might upgrade to a 1080 ti once it goes below 600$.

[reptileexperts]

1 minute ago, EOZ said:

GTX 1080Ti is better deal than RTX 2080

not when DLSS is considered. Also a 1080ti Strix, was the exact same cost new as my 2080 reference finding the best deal on both at 719.00

[EOZ]

1 minute ago, reptileexperts said:

not when DLSS is considered

In which games now you can use it ? 

[reptileexperts]

Just now, EOZ said:

In which games now you can use it ? 

Give it 1 month time. We wont talk about Ray-tracing though... as that will simply destroy a systems performance. 

[DocSwag]

Just now, reptileexperts said:

Give it 1 month time. We wont talk about Ray-tracing though... as that will simply destroy a systems performance. 

Nvidia has to train the neural networks themselves. It takes a lot of time, and they have to do it on a game by game basis. Even in a month I can't see more than 1 or 2 games getting support for it.

 

That being said, $720 is not too much to pay for one. A 1080 ti could be had new for $650 ish but it is a little bit slower.

[Frankenburger]

I was really hoping for an upgrade to Pascal. Historically, every 2 years typically nets a decent jump in performance, even if year 1 of 2 didn't see a product release (eg Maxwell to Pascal). As a Pascal owner, Turing is a disappointment. I wouldn't mind upgrading if I could get at least a 50% bump in performance and the high end cards didn't cost so much. But coming from SLI 1080Ti, it's not really worth it. Also, the fact that the 2080 is the mid range Turing GPU (TU104) and they want $800 for it is extremely off putting. Hopefully the next series cards don't cost as much.

[reptileexperts]

15 minutes ago, Frankenburger said:

I was really hoping for an upgrade to Pascal. Historically, every 2 years typically nets a decent jump in performance, even if year 1 of 2 didn't see a product release (eg Maxwell to Pascal). As a Pascal owner, Turing is a disappointment. I wouldn't mind upgrading if I could get at least a 50% bump in performance and the high end cards didn't cost so much. But coming from SLI 1080Ti, it's not really worth it. Also, the fact that the 2080 is the mid range Turing GPU (TU104) and they want $800 for it is extremely off putting. Hopefully the next series cards don't cost as much.

Definitely the early adopter tax. 

[Frankenburger]

23 minutes ago, reptileexperts said:

Definitely the early adopter tax. 

Indeed. What's even more aggravating about it is Nvidia might be bringing Ray Tracing to Pascal.

 

https://linustechtips.com/main/topic/974470-geforce-41163-driver-ray-tracing-support-for-pascal-with-8gb-vram-or-more/

 

Makes me wonder if high end Pascal will see something comparable to DLSS in the future.

[reptileexperts]

46 minutes ago, Frankenburger said:

Indeed. What's even more aggravating about it is Nvidia might be bringing Ray Tracing to Pascal.

 

https://linustechtips.com/main/topic/974470-geforce-41163-driver-ray-tracing-support-for-pascal-with-8gb-vram-or-more/

 

Makes me wonder if high end Pascal will see something comparable to DLSS in the future.

THAT will make a big difference 

[DocSwag]

2 hours ago, Frankenburger said:

Indeed. What's even more aggravating about it is Nvidia might be bringing Ray Tracing to Pascal.

 

https://linustechtips.com/main/topic/974470-geforce-41163-driver-ray-tracing-support-for-pascal-with-8gb-vram-or-more/

I would assume that it would be much slower than RTX cards at RT though. I guess we'll have to see.

2 hours ago, Frankenburger said:

Makes me wonder if high end Pascal will see something comparable to DLSS in the future.

DLSS works off of the tensor cores in Turing. That means that unfortunately it's not possible for it to come to Pascal; the only thing that might get it is the Titan V

[Frankenburger]

Just now, DocSwag said:

I would assume that it would be much slower than RTX cards at RT though. I guess we'll have to see.

DLSS works off of the tensor cores in Turing. That means that unfortunately it's not possible for it to come to Pascal; the only thing that might get it is the Titan V

I know

It uses predictive image reconstruction to upscale. I don't see any reason why they couldn't eventually do a "dumb approximation", or offer an alternative more akin to checkerboard rendering. I mean, if they can bring Ray Tracing to Pascal on Vulkan via driver updates, then I'm sure they can bring in a comparable upscaling technique to DLSS.

[Lathlaer]

17 hours ago, DocSwag said:

Nvidia has to train the neural networks themselves. It takes a lot of time, and they have to do it on a game by game basis. Even in a month I can't see more than 1 or 2 games getting support for it.

 

Yeah when you talk about implementation in already existing games. If we are talking about future AAA titles then I think Tom Petersen suggested it can be done with the code supplied to them anyway for game ready drivers preparation.

[DocSwag]

20 hours ago, Frankenburger said:

I know

It uses predictive image reconstruction to upscale. I don't see any reason why they couldn't eventually do a "dumb approximation", or offer an alternative more akin to checkerboard rendering. I mean, if they can bring Ray Tracing to Pascal on Vulkan via driver updates, then I'm sure they can bring in a comparable upscaling technique to DLSS.

Ray tracing to Pascal I believe was added much like how ray tracing was added by AMD to their driver stack: they use hardware already there for the computations, but it's much, much slower than on Turing. DLSS explicitly relies on the tensor cores to do its job; if they added an equivalent to older GPUs I would expect it to be much slower than DLSS on Turing, by which point you're probably better off just using normal AA.

 

I'd love to be proven wrong, but sadly I can't see it happening  

[Frankenburger]

5 hours ago, DocSwag said:

I'd love to be proven wrong, but sadly I can't see it happening  

Right or wrong, I'd love to see at least SOME form of upscaling finally hit the PC space and non-Turing cards. DLSS, checkerboard, or something else entirely. Doing so would certainly make a lot more cards "4k worthy", which is pretty important IMHO with 4k 144Hz displays coming out.

[DocSwag]

2 minutes ago, Frankenburger said:

Right or wrong, I'd love to see at least SOME form of upscaling finally hit the PC space and non-Turing cards. DLSS, checkerboard, or something else entirely. Doing so would certainly make a lot more cards "4k worthy", which is pretty important IMHO with 4k 144Hz displays coming out.

AA basically is upscaling  DLSS is supposed to be AA for Turing cards

[Frankenburger]

10 minutes ago, DocSwag said:

AA basically is upscaling  DLSS is supposed to be AA for Turing cards

Downscaling actually

AA renders at a higher resolution than your monitor and resizes to match. Upscaling takes a 1080p/1440p image and approximates it into 4k using things like checkerboard rendering, which is something PC's don't currently have. If you render a 1080p or 1440p image with large amounts of AA, you still get considerable blurring caused by interpolation. Upscaling does a much better job at reducing the blur and provides a much cleaner image.

[DocSwag]

1 minute ago, Frankenburger said:

Downscaling actually

AA renders at a higher resolution than your monitor and resizes to match. Upscaling takes a 1080p/1440p image and approximates it into 4k using things like checkerboard rendering, which is something PC's don't currently have. If you render a 1080p or 1440p image with large amounts of AA, you still get considerable blurring caused by interpolation. Upscaling does a much better job at reducing the blur and provides a much cleaner image.

Whoops I misunderstood you and confused the words .

 

My understanding though is that DLSS behaves similarly to AA, aka it downscales

[Frankenburger]

Just now, DocSwag said:

Whoops I misunderstood you and confused the words .

 

My understanding though is that DLSS behaves similarly to AA, aka it downscales

No worries

 

DLSS is actually more of an AA by upscaling technique, so you're not entirely wrong. Kinda weird in theory considering what we've had access to with past hardware, but Nvidia "trains the AI cores" at their HQ to "reconstruct a supersampled image" from a lower resolution source image. DLSS basically emulates supersampling then upscales. The result is you get 1440p performance at 4k while offering the same or better IQ than 4k with TAA.

 

 

[DocSwag]

29 minutes ago, Frankenburger said:

No worries

 

DLSS is actually more of an AA by upscaling technique, so you're not entirely wrong. Kinda weird in theory considering what we've had access to with past hardware, but Nvidia "trains the AI cores" at their HQ to "reconstruct a supersampled image" from a lower resolution source image. DLSS basically emulates supersampling then upscales. The result is you get 1440p performance at 4k while offering the same or better IQ than 4k with TAA.

 

 

Just read up and you're right, it seems to be a hybrid of sorts. That's actually quite interesting. DLSS seems to be cooler than I thought.

[Frankenburger]

3 minutes ago, DocSwag said:

Just read up and you're right, it seems to be a hybrid of sorts. That's actually quite interesting. DLSS seems to be cooler than I thought.

Indeed. I just hope the task of training the cores for each individual game isn't too daunting of a task for Nvidia. I mean, they can barely keep up with managing SLI, even though most SLI fixes only need a change in compatibility bits. I'd imagine training the AI would take a lot longer than that to get DLSS working properly.

[TheNaitsyrk]

On 22/09/2018 at 4:11 PM, reptileexperts said:

not when DLSS is considered. Also a 1080ti Strix, was the exact same cost new as my 2080 reference finding the best deal on both at 719.00

I'm happy to wait until things get released and then sell my 1080 Ti SLI. Until then I'm happy with what I have. I saturate my displays with 1080 TI SLI far beyond anyway.

[Princess Luna]

On 9/22/2018 at 12:13 PM, reptileexperts said:

Give it 1 month time. We wont talk about Ray-tracing though... as that will simply destroy a systems performance. 

You understand DLSS is just anti-aliasing right? not magic.

 

I mean you made a huge thread to say something every one in here could have told you beforehand, yes a 1070 2waySLI is about the performance of a 2080 when scaling is nice.

 

I understand getting the 2080 Ti since you'd have better experience with a single card and have room for an even better SLI configuration afterwards, returning it for the 2080 was rather a weird decision to me personally.

[reptileexperts]

1 hour ago, Princess Cadence said:

You understand DLSS is just anti-aliasing right? not magic.

 

I mean you made a huge thread to say something every one in here could have told you beforehand, yes a 1070 2waySLI is about the performance of a 2080 when scaling is nice.

 

I understand getting the 2080 Ti since you'd have better experience with a single card and have room for an even better SLI configuration afterwards, returning it for the 2080 was rather a weird decision to me personally.

How so? it's trading SLI what ifs for single card solution is if it performs at 100% all the time plus the ability to go back to SLI and get even better performance. To each their own. Just giving an example out there for folks to reference.