• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.
  • The forums have been upgraded with support for dark mode. By default it will follow the setting on your system/browser. You may override it by scrolling to the end of the page and clicking the gears icon.

Real Performance of the RTX 50 Series (Without DLSS 4, According to Nvidia)

Joined
Oct 10, 2018
Messages
157 (0.06/day)
Today, Nvidia updated their pages with new performance data. The results for the 80/70 series are underwhelming, but I’ve included additional performance data from Far Cry 6 and A Plague Tale: Requiem for context.
(Left to Right: Resident Evil 4, Horizon Forbidden West, Far Cry 6, A Plague Tale: Requiem)


RTX 5070 vs RTX 4070
5070vs4070-2.jpg
5070vs4070.jpg


RTX 5070 = RTX 4070+29% (With performance data averaged over four games.)

RTX 5070 Ti vs RTX 4070 Ti

5070tivs4070ti-2.jpg
5070tivs4070ti.jpg


RTX 5070 Ti = RTX 4070 Ti+29.5% (With performance data averaged over four games.)

RTX 5080 vs RTX 4080

5080vs4080-2.jpg
5080vs4080.jpg


RTX 5080 = RTX 4080+24% (With the average of 4 games)

RTX 5090 vs RTX 4090

5090vs4090-2.jpg
5090vs4090.jpg


RTX 5090 = RTX 4090+34% (With performance data averaged over four games.)

relative-performance-3840-2160-5080-5090.png
relative-performance-2560-1440-5070-5070Ti.png


Overall, we observe greater performance differences in games when DLSS (upscaling) and ray tracing (RT) are enabled simultaneously. However, if Nvidia used more polished data, the average performance could be up to 5% lower in such cases

nvidia-geforce-rtx-5070-performance-chart.jpg
RTX-5070-1.png

4070vs3070ti.jpg


 
Fun fact, there is not a SINGLE set of bars up there that just shows us non-RT raster performance.

If they just boosted the RT part of the config, you might be looking at a complete standstill and perhaps 10-15% from clocks alone. The absence of those numbers seems to be pointing in that direction, if you ask me.

Also lol. 'Full RT', what the AF is that? Some weird mix between it could be PT, could be not?
 
Fun fact, there is not a SINGLE set of bars up there that just shows us non-RT raster performance.

If they just boosted the RT part of the config, you might be looking at a complete standstill and perhaps 10-15% from clocks alone. The absence of those numbers seems to be pointing in that direction, if you ask me.

Also lol. 'Full RT', what the AF is that? Some weird mix between it could be PT, could be not?
Full RT was defined when CP2077 Phantom Liberty Path Tracing update released in 2023.

 
Full RT was defined when CP2077 Phantom Liberty Path Tracing update released in 2023.

Cute. Until nvidia launches another bullshit upsell term. Pretty cool when you get to decide what terms mean and then use them to mislead your audience full time. So we have how many Nvidia layers on top of DX12 ultimate now? Keep in mind they need to sell 6000 series, too, not too long from now, let's see what they've introduced by then.
 
the 5080 it's not great, the others are ok, and if MFG can help even better, i just hope prices come down
 
You can see from the specs there's little uplift other than for the 5090. So it shouldn't surprise.

Normal rendering is conquered pretty much. No reason to upgrade just for that. Which is why Nvidia push RT and upscaling.

The rapid generational progress between generations from 2000-2022 looks pretty over.
 
Cute. Until nvidia launches another bullshit upsell term. Pretty cool when you get to decide what terms mean and then use them to mislead your audience full time. So we have how many Nvidia layers on top of DX12 ultimate now? Keep in mind they need to sell 6000 series, too, not too long from now, let's see what they've introduced by then.

It's always the same song and dance. New products are coming and all of a sudden all your current stuff is no longer good. Nvidia is more than happy to show you the artifacts introduced by DLSS when it comes to selling you new cards. Wash, rinse, and repeat.
 
Full RT was defined when CP2077 Phantom Liberty Path Tracing update released in 2023.


Better to wait for real results if you have the money or not. Either way not interested in buying a GPU with a rubber seal as seals tend to fail sooner or later and as seen as were talking liquid metal it's a big nope.
 
Last edited by a moderator:
Want full-on real-time RT, just turn away from whatever device you're using to access interweb and witness real life. At whatever FPS your eyes and brain can ever process.

What we see in games is just a demo version for a demo version, everything is blurred, slurred and distorted, it's truly far, far away from real life lighting and movement. The Brainporium logo is still broken in CP77 btw, it clearly shows up red and static in that reflection despite changing its colours and fill.

Not to devalue any of that, it's just not to be called FULL RT by any mean. This is a cool feature and I'm gonna be happy to see more and more progress in this regard, just... The marketing is a tad too deceptive. Even for a monopolist.

Speaking non-RT performance, we clearly don't need much more raster speed than we already have. Of course more is better but standard graphics can run on a 500 dollar GPU just fine even at 4K, it's RT that pushes GPUs to their limits and overpowers them. And badly optimised game engines such as UE5, CryEngine and the likes.

That said, RTX 50 series is exactly what we could expect. Much more AI tuned, moderately more RT oriented and stagnant at pure raster. Nobody should expect serious improvements over Ada in rasterised activities anyway.
 
So basically it is all rumor and suspicion based on vague 1st party information. These sorts of threads should be marked accordingly. Until there are reviews from independent 3rd parties, it is all FUD.
 
Just reduce the stated numbers by a few % for obvious reasons and you have 5070 = 7900GRE in raster.

So much for 5070 = 4090 claims and all those who claimed nvidia will have 50% generational uplift or whatever. I mean nvidia have done this for a few generations now by showing results that aren't directly comparable and misleading at best just so they can claim their 3/4x performance increases. Standard stuff.

5090 looks like the only interesting one of the bunch simply because that massive bandwidth lends well to non-gaming workloads and should also help remove 4090's bandwidth bottleneck entirely.
 
soooo..... 5070=4070Ti, same perf, same vram, almost same power, more AI frames
5070Ti=4080/S,same perf, same vram, almost same power, more AI frames

this gen is a hard pass, especially when 4000 series owners still get the updated dlss SR and RR, just not the Multi Frame FG
9070xt is gonna sell really well.
 
this gen is a hard pass, especially when 4000 series owners still get the updated dlss SR and RR, just not the Multi Frame FG
9070xt is gonna sell really well.
I think that's kind of many peoples sentiments it seems. If you have anything above or at a 4060ti you can probably reasonably pass on this gen tbh. The 9070 and 9070XT could do very well now that NVIDIA have opened themselves up. I'm kinda banking on that fact. Will wait and see though, I might just jump on the 9070XT and finally buy a new GPU for once instead of a used one for my next upgrade lol.
 
The reason for the RT only is because Nvidia removed a bottleneck inside of the architecture that decreased the latency inside the pipeline of the RT rending speed. Memory bandwidth probably also slightly helps, but it probably will do little to nothing for rasterization part of the rendering. It's still bottlenecked by being tripled linked to shaders (170 SM's), texture units (680) & Rt/tensor cores (170) that still have to fight over having only 192 Rend out puts (R.O.P's) compared the total amount RT cores. (Also, when did the R.O.P's go down from 224 to 192 on RTX 5090?) Do not expect anything in rasterization other than maybe 20%- to 25% on average.
 
Last edited:
i think this generation is fine, the problem is people that want to buy a new shiny gpu every generation, that's what is idiotic, you really can't blame Nvidia if you act like this, if you want to be milked by every penny they would be stupid to say no. When you milk a cow you try to accommodate their needs.
I skipped a generation and i think the possible gains to a 5070 are just fine (i just need a 16gb version) or possible a 5060ti
 
The reason for the RT only is because Nvidia removed a bottleneck inside of the architecture that decreased the latency inside the pipeline of the RT rending speed. Memory bandwidth probably also slightly helps, but it probably will do little to nothing for rasterization part of the rendering. It's still bottlenecked by being tripled linked to shaders (170 SM's), texture units (680) & Rt/tensor cores (170) that still have to fight over having only 192 Rend out puts (R.O.P's) compared the total amount RT cores. (Also, when did the R.O.P's go down from 224 to 192 on RTX 5090?) Do not expect anything in rasterization other than maybe 20%- to 25% on average.

What about heavy RT/PT (without FG)?
I don't think raster performance has to improve when we already achieve 400-500fps.
I want good RT/PT before any DLSS and good DLSS without FG.
 
What about heavy RT/PT (without FG)?
I don't think raster performance has to improve when we already achieve 400-500fps.
I want good RT/PT before any DLSS and good DLSS without FG.

You won't get better RT or PT unless they remove them from being linked together with all the other rasterization parts. If you look at Nvidia's desgins & both AMD all almost all increase of RT/Pt have come from the massive increases in rasterization. If the improvements were just straight up RT/PT then there is no need to increase in render out puts (R.O.P's) But since these cards are actually hybird cards that a mix of rasterization & raytracing they'll always have limitation.s
 
Fun fact, there is not a SINGLE set of bars up there that just shows us non-RT raster performance.

The absence of those numbers seems to be pointing in that direction, if you ask me.
Yep, Nvidia have historically been happy to brag about >20% generational increases, even if it's just clockspeed bumps and barely an IPC improvement.
The fact that non-RT performance numbers are glaringly absent means we can probably expect no improvement at all, beyond core count and clock increases over last gen models.
 
You won't get better RT or PT unless they remove them from being linked together with all the other rasterization parts. If you look at Nvidia's desgins & both AMD all almost all increase of RT/Pt have come from the massive increases in rasterization. If the improvements were just straight up RT/PT then there is no need to increase in render out puts (R.O.P's) But since these cards are actually hybird cards that a mix of rasterization & raytracing they'll always have limitation.s
rt still runs on fp32/int32, but do you, by any chance, know what "other graphics" is on that chart though ?

236365_RTX_Frame.jpg
 
rt still runs on fp32/int32, but do you, by any chance, know what "other graphics" is on that chart though ?

236365_RTX_Frame.jpg
Don't believe anything Nvidia says about RT. They want you to think that it's a single one they use, since their better at, but it's not. FP32 is small part of that.
It isn't just floating points calculations as their texture calculation inside of that, that are part of the render & reshading for RT color depth. ( Like for shadows & light off textures)
 
rt still runs on fp32/int32, but do you, by any chance, know what "other graphics" is on that chart though ?

Pass, only Nvidia will tell you the answer for sure, but there's a bunch of graphics-related operations that don't run on the INT shader cores, stuff like (quoting the 40-series architectural overview):

Shader Execution Reordering (SER)
Displaced Micro-Meshes
Opacity Micro-Masks
FP8 inferencing

It sounds like those are related to shader operations but unlikely to run on the FP32, INT32, or RT cores.
 
1737642399606.png


So this is proof of Nvidia's claims. 5090 is 35% faster than 4090 in 4K. RTX 5070/Ti could be the best GPUs, along with 5060/Ti, in this generation. There is a price drop of $50 and a 1.3 performance uplift compared to the 40 series. The RTX 5060 Ti 16GB would compete against RTX 4070, even 4070 Super, if it has more cores than 4608. On the other hand, RTX 5060 would have the same level of performance as the 3070 Ti (I hope it will come with 12GB but 128 bit).
 
So this is proof of Nvidia's claims. 5090 is 35% faster than 4090 in 4K. RTX 5070/Ti could be the best GPUs, along with 5060/Ti, in this generation. There is a price drop of $50 and a 1.3 performance uplift compared to the 40 series. The RTX 5060 Ti 16GB would compete against RTX 4070, even 4070 Super, if it has more cores than 4608. On the other hand, RTX 5060 would have the same level of performance as the 3070 Ti (I hope it will come with 12GB but 128 bit).
To me it looks like IPC is near-zero.

RTX 5090 has 33% more cores, and uses 33% more power to go 35% faster.

The 5090 is the only GPU that's had double-digit percentile core increases. The 5070 is a significant core count reduction from the 4070S so it may even perform worse in some games without the help of DLSS4 and MFG.
 
Back
Top