I have noticed that FPS does not seem to be scaling linearly either with GPU overclocking or with differences between the 3080 and 3090.
 
For example, a 5% overclock does not seem to yield a 5% fps increase across either card.
 
The impact of sub-system "stuff" is not linear.  
For example, 3080 vs 3090 (Data taken from videocardz.net)
3090 has about 20% more "stuff" but gaming benchmarks yield 5-10% improvements to FPS
                      3090         3080    3090 delta
CUDAs            10496        8704   +20%
Tensor Cores   328              272   +40%

RT Cores         82                 68   +20%

ROPs             112                96    +16%
 
Since these things are electronic machines there has to be an explanation somewhere.  A driver bottleneck, an i/o bottleneck, throttling.. odd wait states ..something..
I have not seen a plausible technical explanation.  Has anyone got technical insight into this?
 
EDIT:  To reduce complexity, I am only considering the impact of overclocking on 4K gaming. 

• Note the impact of CPU bottlenecks at 1440 or less, is well understood. 
• The impact of textures on 8K gaming and the need for a 3090 at this resolution is also understood.

eknight76
While this doesn't exactly answer your question, this article may shed some light:
www.techspot.com/review/2110-nvidia-rtx-2080-1440p-gaming-bottleneck
 
Scaling appears to differ across resolutions and, according to the article, it appears largely due to the type of shaders being utilized for the workload.  So likely, the larger number of cores and/or shaders does not aid in all workloads equally.  3090s really shine only in the highest of resolutions, even up to 8k. 
 
Edit:  Not use to this message board software, for some reason, my hyperlink wasn't showing up.

Thank you
This is a good article and focuses on the lack of scaling between resolutions.  It provides data on fps at different resolutions, and identifies where to expect CPU bottlenecks.  I believe this is well understood by the community.
 
What the community seems to be missing is an answer to: Why when playing @ 4K if I overclock by 10% don't I get a 10% increase in framerate or if I use a 3090 with 20% more stuff, why don't I get 20% more FPS.
 

Diminishing returns maybe? internal card bottlenecks we can't measure maybe. Personally I don't think it being electronic machines really says that much. If you are measuring just the electrons moving through the card maybe you'd get 20% for 20% more of something, but it's not that simple. These cards are running programming code instructions that themselves aren't 100% efficient. I'd imagine that the execution of code has diminishing returns in its own right. These are all just guesses, but I don't think it's logical to assume the different parts of gpu hardware are all perfectly efficient and will provide linear boosts like that. Also with higher "stuff" you get higher heat output (because they aren't efficient and all the electricity that goes in a portion of that is going to come out as heat instead of being used for work) which means worse stability. 

When, in the history of computing, has "clock speed" scale directly equaled FPS in video games? Not being a jerk, but there are a lot of variables (many of which you named). So I feel like you answered your own question.

To me that's like saying a semi-truck with 700 whp should be 16% faster in the quarter mile than a Corvette Z06 with 600 whp. And it's like, no, hp doesn't directly scale over performance in the first place (you get diminishing returns against drag; I believe it was Bugatti that said after 200 MPH, every 1 MPH increase in top speed is a 100 whp requirement roughly just to overcome the immense amount of aerodynamic drag), not to mention you are comparing different power bands and torque curves, different gear ratios and weights, different wheel sizes and weights, different amounts of traction, different drive wheel configurations, different vehicle weights, different vehicle aerodynamic efficiencies, different-

Yeah, you get the idea. lol. You seem pretty smart, I don't know why you would think it has a direct, 100% 1:1 correlation to frame rate on every piece of software ever made.

riku98523
Diminishing returns maybe? internal card bottlenecks we can't measure maybe. Personally I don't think it being electronic machines really says that much.  If you are measuring just the electrons moving through the card maybe you'd get 20% for 20% more of something, but it's not that simple. These cards are running programming code instructions that themselves aren't 100% efficient. I'd imagine that the execution of code has diminishing returns in its own right.
 
 
The above is basically true. In a sense this is a restatement of the question I asked.  We are not talking about moving electrons.  It is better to think we are changing states through mathematical transforms as we move through the processing pipeline.  Theoretically the pipeline can be optimized so that it scales with clocks.  However, if elements of the pipeline are mismatched you can run into queue overruns, wait states etc.  The propensity for these problems is what I am trying to understand through a technical discussion that would include a view of the architecture and how the components interact.
 
These are all just guesses, but I don't think it's logical to assume the different parts of gpu hardware are all perfectly efficient and will provide linear boosts like that.  Yes, this is what I want to find out specifically.
 
Also with higher "stuff" you get higher heat output (because they aren't efficient and all the electricity that goes in a portion of that is going to come out as heat instead of being used for work) which means worse stability. 
Heat and stability are real issues but can be ignored for the purposes of this analysis, becasue they can be mitigated through external means or simply by reducing system stress.

Thanks for the input though

irakandjii
I have noticed that FPS does not seem to be scaling linearly either with GPU overclocking or with differences between the 3080 and 3090.
 
For example, a 5% overclock does not seem to yield a 5% fps increase across either card.
 
The impact of sub-system "stuff" is not linear.  
For example, 3080 vs 3090 (Data taken from videocardz.net)
3090 has about 20% more "stuff" but gaming benchmarks yield 5-10% improvements to FPS
                      3090         3080    3090 delta
CUDAs            10496        8704   +20%
Tensor Cores   328              272   +40%

RT Cores         82                 68   +20%

ROPs             112                96    +16%

Since these things are electronic machines there has to be an explanation somewhere.  A driver bottleneck, an i/o bottleneck, throttling.. odd wait states ..something..
I have not seen a plausible technical explanation.  Has anyone got technical insight into this?
 
EDIT:  To reduce complexity, I am only considering the impact of overclocking on 4K gaming. 

• Note the impact of CPU bottlenecks at 1440 or less, is well understood. 
• The impact of textures on 8K gaming and the need for a 3090 at this resolution is also understood.

It's mainly due to power limit.

kevinc313
 
 
It's mainly due to power limit.

Can you elaborate or provide a technical reference?  I am interested in your reasoning.    
 
Power is generally a function of frequency and voltage.  At a fixed voltage, power dissipated increases with frequency.  So it will limit the max frequency you can reach at a given voltage, but generally would not impact the relationship of fps to overclock, just max overclock you can reach.

irakandjii

arestavo
Hardware unboxed goes a little bit into it as to why 4K scales better than 1440P (or 1080): https://youtu.be/c7npXXyXJzI

Interesting this video contains almost exactly verbatim comments from the Techspot.com review mentioned above, I strongly suspect one is plagiarizing the other. 
 
It again focuses on the impact of changing resolution not using 4K and overclocking or upgrading to a 3090.
 
 

LMFAO, I thought you were joking at first but seems serious. 
 
Techspot = The written form of Hardware Unboxed.

eknight76
That all said, irakandjii does bring up a good consideration, the 3090 scales unusually bad compared to prior generations for being a ti type upgrade.  From my understanding, tensor cores do not help much with rasterization, so the 40% uptick there is probably doing much for frames.

The Tensor cores are used in AI programing for deep learning, so I suspect they will drive improvements to DLSS.  My guess is as more games adopt DLSS they will have a significant impact. 
 
I just don't know where they fit in the pipeline and if they inflict penalties (delays) on workloads that don't have DLSS.  Are they always activated or stepped over...  humm
 
Interesting, a thread to pull on this ball of twisted twine.

irakandjii

USALion
When, in the history of computing, has "clock speed" scale directly equaled FPS in video games? Not being a jerk, but there are a lot of variables (many of which you named). So I feel like you answered your own question.
 
There are many cases where there was linear scaling and I am not expecting perfect 1 to 1 scaling, more so when we were dealing with pure rasterization.  
 
To me that's like saying a semi-truck with 700 whp should be 16% faster in the quarter mile than a Corvette Z06 with 600 whp.
No, acceleration is a function of mass and force.  The useful engineering metric is power to weight ratio.
 
And it's like, no, hp doesn't directly scale over performance in the first place (you get diminishing returns against drag; I believe it was Bugatti that said after 200 MPH, every 1 MPH increase in top speed is a 100 whp requirement roughly just to overcome the immense amount of aerodynamic drag)
Power lost to drag is a function of Velocity cubed, it is not a diminishing rate of return but a direct function approaching a limit.  Also drag is basically irrelevant below 55 mph.  It is why the USA designated 55 mph during the energy crisis.  At the same time, drag is severely limiting at high speeds as is rolling resistance. 
This kind of explanation describing a very complex relationship is what I am looking for with FPS scaling on the 3080 /3090.
 
, not to mention you are comparing different power bands and torque curves, different gear ratios and weights, different wheel sizes and weights, different amounts of traction, different drive wheel configurations, different vehicle weights, different vehicle aerodynamic efficiencies, different- 
 
All of these factors can be evaluated with a series of linear equations.  Many of these engineering parameters are design inputs and established for specific use cases for the vehicle at conception.  I can point you to a good engineering reference on Amazon:  Theory of Ground Vehicles Hardcover – Illustrated, Aug. 4 2008 , 
 Author J. Y. Wong.  I took his course as an elective in my 4th year.

Yeah, you get the idea. lol. You seem pretty smart, I don't know why you would think it has a direct, 100% 1:1 correlation to frame rate on every piece of software ever made.
 
Not 1:1 but understanding why it is so badly off would be useful as stated above.
Just because there are a lot of variable does not mean we should stop trying to understand.  In part I am trying to determine if there are possible driver changes that can significantly improve the effectiveness of Overclocking or the 3090.  If so it makes sense to buy a strong "over clocker" or even a 3090 instead of a 3080. 
 
If not then accept what is and buy a cheap 3080 and wait for the next generation.

Thanks for the input I hope this better describes what I am looking for.

Great feedback and perspective from an engineering and physics standpoint. I didn't mean to imply the performance of the vehicle (and video card) were impossible to calculate, I just wouldn't consider only clock speed or cores to directly correlate to an exact percentage increase in performance due to all of the variables.. some of them software, both driver and the game in question's code.

irakandjii

kevinc313
 
 
It's mainly due to power limit.

Can you elaborate or provide a technical reference?  I am interested in your reasoning.    
 
Power is generally a function of frequency and voltage.  At a fixed voltage, power dissipated increases with frequency.  So it will limit the max frequency you can reach at a given voltage, but generally would not impact the relationship of fps to overclock, just max overclock you can reach.

When comparing the two cards, look at their power limits vs. core count vs. memory vs. component power draw (GPU-Z).