Choosing between 2 GPUs - help on 2 basic principles needed

Coolio

Weaksauce
Joined
Jan 8, 2021
Messages
118
Hi folks,

I will have to compare Nvidia GPUs in different variations soon (Reference, FE, AIB) and will appreciate your help with some principles:

  1. I've always thought that at the end of the day the most significant tech specs become part of certain formulas and finally result in 2 parameters: flops (how fast the GPU performs calculations) and bandwidth/transfer rate (how much data it can push to the system at a time). So I supposed I can use those 2 to compare GPUs of at least the same brand and generation, for example Nvidia RTX 2060 Founders Edition and some ASUS AIB model built on the same GPU. However, someone has pointed to the example of AMD Radeon Vega 64 and Nvidia GeForce RTX 2070. Yeah, these are different brands, but still - Vega has 11518 Gflops and 483.8 Gb/sec bandwidth and GeForce has 8920.32 Gflops and 448 Gb/sec, however both cards are relatively competitive. So finally, can I use flops/bandwidth to compare GPUs and if yes - to which extent?
  2. Which scenario will you choose for better performance and why: 1) "less clocks & very fast VRAM (HBM2)" or 2) "higher clocks & fast VRAM (GDDR5X)"?
Tank you for your comments!
 
^^ that.
And if you are still hell bent on details, look at benchmarks in games you actually intend to play. Not all will be the same.
 
Thank you guys!

What embarasses me is that GPU seems to be the only PC component which can't be compared using its tech spec.
You can compare RAM, SSD, CPU, PSU, monitor - from their performance perspective, but when it comes to GPU the comparison is boiled down to just watching YouTube videos with side-by-side comparison in games.

By no means I'm underestimating the benefits of the "visual" part of the evaluation, but doesn't it seem the same odd to you as the proposal to "copy 50Gb of data to see which SSD works for you best"? To me - it definitely does.
 
Last edited:
Think of it like cars. You have a c8 corvette and a hellcat. on paper a hellcat is 700hp and c8 495. Yet from a dig the c8 will win almost everytime while from a roll the hellcat will win. The c8 will carve corners and track faster too.

The same with gpu. Need to look at where you need to apply that performance and go with the one that has strengths in the ones you use.
 
Thank you guys!

What embarasses me is that GPU seems to be the only PC component which can't be compared using its tech spec.
You can compare RAM, SSD, CPU, PSU, monitor - from their performance perspective, but when it comes to GPU the comparison is boiled down to just watching YouTube videos with side-by-side comparison in games.

By no means I'm underestimating the benefits of the "visual" part of the evaluation, but doesn't it seem the same odd to you as the proposal to "copy 50Gb of data to see which SSD works for you best"? To me - it definitely does.
I don't think CPU performance is as easily compared either. Look at clock speeds from the Pentium 4 days until now and you don't really have major changes, but IPC, core count, etc. has ensured higher performance today even at similar clock speeds. RAM is probably the one thing I can think of that's relatively straightforward to compare (until you get into tuning it). PSUs you can compare watts, but there's still good vs bad PSU brands at the same wattages, SSDs you can compare storage speed but there's multiple other factors, and monitors are a hell to compare unless you just care about size and price.
 
  • Like
Reactions: Dan_D
like this
Think of it like cars. You have a c8 corvette and a hellcat. on paper a hellcat is 700hp and c8 495. Yet from a dig the c8 will win almost everytime while from a roll the hellcat will win. The c8 will carve corners and track faster too.

The same with gpu. Need to look at where you need to apply that performance and go with the one that has strengths in the ones you use.
Dammit - you managed to find the comparison hard to argue with! :))))) The problem is you can test-drive the car before you buy it. But ok, I hear you and looks like I have no choice but take it as it is: only side-by-side comparison works with GPUs. Sad but true.


I don't think CPU performance is as easily compared either. Look at clock speeds from the Pentium 4 days until now and you don't really have major changes, but IPC, core count, etc. has ensured higher performance today even at similar clock speeds. RAM is probably the one thing I can think of that's relatively straightforward to compare (until you get into tuning it). PSUs you can compare watts, but there's still good vs bad PSU brands at the same wattages, SSDs you can compare storage speed but there's multiple other factors, and monitors are a hell to compare unless you just care about size and price.
Yeah, don't take me wrong - I'm not trying to oversimplify things, but it has always seemed to me it's possible to compare all the above-mentioned HW based on its tech spec - at least in wide strokes. And even CPUs - we can be more or less precise at least within the same brand's lineup (even more so - within the same architecture). OK, looks like GPUs are the ethereal world to just accept and save nerves. :)))
 
  • Like
Reactions: x509
like this
Muppets rock :) I think the problem with any particular spec that is published is you are not really getting an apples to apples comparison, as not only are you comparing a corvette to a hellcat, you are also getting a completely re-designed car generation to generation.
Also the kind of track you are driving on does make a difference is where you are getting gains, as the tracks tend to also get more complex to navigate. Basically A+ games get more complex, AI nural simulations get more complex, video encoding add new codecs, luckily dx12 stays etc.
Dont forget to factor in new gens of gpu architectures add cheats to make tasks easier (like compressing data over pipelines basically why nvidia gpu can out perform amd in certian tasks even though on paper better specs) like a turbo. So best bet unless you are using a gpu to do calculation stuff is really do a suite of benches and try to compare as close to what you have to see if it will get you there.
 
Because raw math of one kind (what Gflops represents) doesn't directly apply to a specific task (rendering a frame of a game, mining a coin, etc) equally, depending on WAY too many factors. Modern GPUs have 40% more transistors than a top-end consumer CPU these days, never mind all the other associated coprocessors, RAM, etc on them. It's basically a complete system in and of itself - and they're each optimized slightly differently for different tasks, much like a workstation system with a ThreadRipper Pro is different than an overclocked Z490+10900K... different tasks, different uses, different outcomes. HP isn't everything in the car analogy - one type of math isn't everything in this. :)
 
lopoetve Thank you mate! If this thread doesn't persuade someone to stop comparing GPUs by tech spec - nothing will. 😆
 
If you already decided a model and are now just wondering if you should get an FE, or AIB card, it comes down to really just the benchmarks, but also which one will keep the card cool, and of course which will fit in your case if you have any size/space issues.
 
If you already decided a model and are now just wondering if you should get an FE, or AIB card, it comes down to really just the benchmarks, but also which one will keep the card cool, and of course which will fit in your case if you have any size/space issues.
So the difference between Reference/FE/AIB of the same GPU is mainly in cooling, right? Or may other specs (VRAM size at the very least, smth. else?) differ as well?
 
Most common specs will not differ. Ram size can but its pretty rare. The components used to build the PCB such as the number and type of VRM are often vary different but this may not mean much if you are leaving the cards at stock clocks. If you were planning to overclock then you would typically look to see what AIB is using a more "overbuilt" VRM / power delivery on their card. AIB's can sometimes up the clocks memory speed and power targets above reference based on this themselves (think OC models)
 
If you were planning to overclock then you would typically look to see what AIB is using....
AIB's can sometimes up the clocks memory speed and power targets above reference
No OC in plans for sure. I just want my card to be stable and its capabilities to not be limited by lousy self-cooling. So my choice will be between Reference and FE I guess.
Btw, are binned chips (which FE is supposed to be built on, right?) really more efficient than non-binned ones (which are used in Reference edition)?

I guess the only reason to go for AIB will be if I use some really powerful card like GeForce RTX 3070 (220W), because my case form factor will be SFF sub-10L: Dan A4 (7.2L) or Louqe Ghost S1 (8.2L) or FormD T1 (9.5L). However, I suppose if I use say GeForce GTX 1070 (150W), then its standard (Reference) cooling will be enough even in small Dan A4, am I right?
 
  • Like
Reactions: x509
like this
I'm pretty sure Optimum Tech tested the ref design in an SFF case (the T1 I believe).

I would avoid reference in a sandwich layout, personally.

Edit: oh, 3070. Does that cooler blow through like the 3080s and 3090s?
 
I would avoid reference in a sandwich layout, personally.
Sandwitch means any build where cards are placed horizontally one over another? Isn't it 90% of builds and why particularly Reference edition GPU sucks in this scenario?

Edit: oh, 3070. Does that cooler blow through like the 3080s and 3090s?
Not sure I got what you meant. I don't have either yet, so can't compare their cooling capabilities if you meant that...
 
Think of it like cars. You have a c8 corvette and a hellcat. on paper a hellcat is 700hp and c8 495. Yet from a dig the c8 will win almost everytime while from a roll the hellcat will win. The c8 will carve corners and track faster too.

The same with gpu. Need to look at where you need to apply that performance and go with the one that has strengths in the ones you use.
Ya you helped :ROFLMAO: :ROFLMAO: :ROFLMAO:
 
Sandwitch means any build where cards are placed horizontally one over another? Isn't it 90% of builds and why particularly Reference edition GPU sucks in this scenario?


Not sure I got what you meant. I don't have either yet, so can't compare their cooling capabilities if you meant that...
Sandwich refers to the layout of the case. Dan A1, Formd T1 and Ghost S1 are all sandwich layouts. Which means you have the GPU on one side and the mobo, cpu etc on the other.
 
Sandwich refers to the layout of the case. Dan A1, Formd T1 and Ghost S1 are all sandwich layouts. Which means you have the GPU on one side and the mobo, cpu etc on the other.
OK and why exactly Reference type is bad for this layout, while FE and AIB are OK?
 
So the difference between Reference/FE/AIB of the same GPU is mainly in cooling, right? Or may other specs (VRAM size at the very least, smth. else?) differ as well?
Difference in cooling solution yes, but might also have some different clock speeds (marginally different - AIB might have a slight overclock out of the box vs FE), and non-reference cards might have different power delivery. They will also have a completely different PCB design from the FE card. But essentially a 3070 is a 3070 in terms of what you are getting. So just figure out what works for your case/needs in terms of size, cooling solution, price, noise, etc.

Honestly I think you are really overthinking this.
Edit: oh, 3070. Does that cooler blow through like the 3080s and 3090s?
Yes it does.
 
Last edited:
I'm pretty sure Optimum Tech tested the ref design in an SFF case (the T1 I believe).

I would avoid reference in a sandwich layout, personally.

Edit: oh, 3070. Does that cooler blow through like the 3080s and 3090s?
When you say reference, do you mean the Founder's Edition which is NOT the reference design this gen?
 
non-reference cards might have different power delivery. They will also have a completely different PCB design from the FE card.
The difference in PCB is in favor of (future) OC? So no specific benefit from this difference if I don't go OC way, right?
Btw, are binned chips (used in FE as I know) worth its money compared to standard ones used in Reference, or is it mostly marketing?

Honestly I think you are really overthinking this.
Not only, I'm overconfused too :D taken that:
When you say reference, do you mean the Founder's Edition which is NOT the reference design this gen?
Doesn't FE of every gen have NOT a Reference design? I'm confused.. Which type (and why) is finally NOT recommended (as Sir Psycho said) for those 3 sub-10L cases I've mentioned?
 
The difference in PCB is in favor of (future) OC? So no specific benefit from this difference if I don't go OC way, right?
Btw, are binned chips (used in FE as I know) worth its money compared to standard ones used in Reference, or is it mostly marketing?


Not only, I'm overconfused too :D taken that:

Doesn't FE of every gen have NOT a Reference design? I'm confused.. Which type (and why) is finally NOT recommended (as Sir Psycho said) for those 3 sub-10L cases I've mentioned?

Usually Founder's Edition cards are considered reference designs in past generations. But that is not true for the 30-series. The Founder's Edition cards for the 3090, 3080, 3070, 3060 Ti are NOT reference design. There are some AIB's that might be using reference designs, but not sure who.

Honestly even the AIB cards have some pass through on their cooling designs like the FE cards. So I am not exactly sure what Sir Psycho is getting at either.
 
Usually Founder's Edition cards are considered reference designs in past generations
And by "design" you mean PCB, right? Not physical design and not difference in # of fans (which obviously follows from the physical design).

So I am not exactly sure what Sir Psycho is getting at either.
He will possibly clarify, which I will appreciate as I've never heard of a particular GPU edition to be not recommended for sub-10L cases on the grounds of their sandwich card layout.

----
P.S. Still looking for advice on:
  1. Are binned chips (used in FE) worth paying for them or not (if I'm not going to OC)? For example there's a Reference card which is OK for me (incl. cooling), but shall its FE version (with binned chip and better cooling) be a better choice (taken it fits my budget)?
  2. Can one say that for sub-10L cases I've mentioned above, from the cooling perspective it makes to start considering FE/AIB only from a certain wattage? In other words GPU below __W is fine in its Reference version and buying its FE/AIB edition will be excess spending?
 
And by "design" you mean PCB, right? Not physical design and not difference in # of fans (which obviously follows from the physical design).


He will possibly clarify, which I will appreciate as I've never heard of a particular GPU edition to be not recommended for sub-10L cases on the grounds of their sandwich card layout.

----
P.S. Still looking for advice on:
  1. Are binned chips (used in FE) worth paying for them or not (if I'm not going to OC)? For example there's a Reference card which is OK for me (incl. cooling), but shall its FE version (with binned chip and better cooling) be a better choice (taken it fits my budget)?
  2. Can one say that for sub-10L cases I've mentioned above, from the cooling perspective it makes to start considering FE/AIB only from a certain wattage? In other words GPU below __W is fine in its Reference version and buying its FE/AIB edition will be excess spending?
if you can find a gpu you want for sale then buy it casue the rest of us have just about lost interest in buying cards constantly sold out or way over original msrp
 
I hope those "rest" have not lost interest in sharing their experience with others though. :)
i just think some things/questions are made to be more complicated than they really are, but finding a nice/new model card in stock for regulare msrp is like winning the lottery these days
 
I LOVE Slade's comments comparing the C8 Corvette vs the Hellcat.

With the scarcity of gpus in the near future, just getting a high end gpu will be fine.
 
Back
Top