NVIDIA GPU Generational Performance Part 1 @ [H]

FrgMstr

FrgMstr

Just Plain Mean
Staff member
Joined
May 18, 1997
Messages
55,596
NVIDIA GPU Generational Performance Part 1

Ever wonder how much performance you are really getting from GPU to GPU upgrade in games? What if we took GPUs from NVIDIA and AMD and compared performance gained from 2013 to 2018? We are going to start that process today in Part 1 focusing on the GeForce GTX 780, GeForce GTX 980 and GeForce GTX 1080 in 14 games.

If you like our content, please support HardOCP on Patreon.
 
lucidrenegade said:
Nice performance jump from the 980 to 1080, compared to the 780 to 980.
Click to expand...

Could be other factors at work, but I would guess that much of that is due to the 1080 having double the VRAM of the 980, versus the 980 only having 1 GB more than the 780.
 
Not sure if it's just me, but I don't see any images. Such as charts, for example.
 

Attachments

  • Screenshot_20180725-105125.png
    Screenshot_20180725-105125.png
    607.5 KB · Views: 0
Last edited by a moderator:
auntjemima said:
Not sure if it's just me, but I don't see any images. Such as charts, for example.
Click to expand...

You have a voicemail.

The charts work on desktop, haven't tried mobile myself. Maybe hold the phone sideways? That shows more information when you're on the forum homepage.
 
That was a good read to think the 780 gtx met it's maker Kingdoms Come. Buying the newest and best card isn't a fools game after all.
 
scojer said:
You have a voicemail.

The charts work on desktop, haven't tried mobile myself. Maybe hold the phone sideways? That shows more information when you're on the forum homepage.
Click to expand...

I never check voicemail lol..

I tried sideways and requested the desktop page. No good.

I'll just read it when I'm home.
 
DigitalGriffin said:
If we scaled the stock clock so it was an apples to apples comparison for IPC purposes based on CU performance. (I know it's not apples to apples entirely because of memory advances)
Click to expand...

Who cares? If the generational improvement is a 40% increase in clock speed with a 10% decrease in IPC, that's still a 26% increase in performance.
 
DigitalGriffin said:
While I really enjoyed this article, there's some data left out I would have liked to see.

If we scaled the stock clock so it was an apples to apples comparison for IPC purposes based on CU performance. (I know it's not apples to apples entirely because of memory advances)

I think we will be stuck with 2GHz for the immediate future, so if there aren't IPC gains, it could be a good indicator of where things are headed design wise.
Click to expand...
No GPU out there runs at 2 GHz out of the box right now, so I don't know what you mean about being "stuck" with 2 GHz. I also don't understand the obsession with figuring out IPC on a GPU when it isn't the most important metric of its performance in the least.
 
Armenius said:
No GPU out there runs at 2 GHz out of the box right now, so I don't know what you mean about being "stuck" with 2 GHz. I also don't understand the obsession with figuring out IPC on a GPU when it isn't the most important metric of its performance in the least.
Click to expand...

Quite simply this:

Look where intel has been suck from i2XXX->i6XXX. They hit a MHz ceiling. I think the same is coming for graphics cards.

Adding more Cuda cores has limitations on improvements.
 
Last edited by a moderator:
DigitalGriffin said:
While I really enjoyed this article, there's some data left out I would have liked to see.

If we scaled the stock clock so it was an apples to apples comparison for IPC purposes based on CU performance. (I know it's not apples to apples entirely because of memory advances)

I think we will be stuck with 2GHz for the immediate future, so if there aren't IPC gains, it could be a good indicator of where things are headed design wise.
Click to expand...

Here's a car analogy. You are a car reviewer, you get 5 different cars to test off-road performance. Sure, you can put your own custom off road tires on each car. However, by doing this you are modifying the default stock performance to your will. The only way to keep things fair, is to test everything at default stock, as the manufacturer intended its performance profile to be.

I get what you are saying about IPC, but we will never, ever, be able to create an apples-to-apples situation because at the end of the day they all have different CUDA Core counts, ROPs, Texture Units, and other dissimilar specs that keep us from achieve that, even if we matched clock speeds. We still wouldn't be really testing anything useful.

The only way to keep things fair is to evaluate the default performance, as the manufacturer intended at launch. That's the performance that really matters because that's the performance you'll get in-game when you buy the card.
 
Nice review. The only thing I'd have liked to see in addition would have been the 70 series, but that's only because I'm using a 970, so bias. :) I get why the focus is on 80 and ti.

I somewhat agree with the question posed near the end of conclusion 3, that except for a handful of games the 1080 hasn't really been taxed much by anything -- at 1440p. Of course, 4K is becoming the new hotness, so it might behoove [H] to conduct the performance review of the 1080 ti at both 1440p and 4K, since it's just going to stomp all over 1440p. 4K results will give a better picture of where we may be heading with 1180 and 1180 ti.
 
Man... 980 to 1080 was like i7 920 to i7 2700k kind of bump. 780 to 980's been what we've seen since from Intel.

Taking bets: 1080 to 1180 will be (as a %) a gain closer to 780 -> 980 rather than 980 -> 1080. Call it a hunch that I don't see lightning striking twice, back to back.
 
Neat article.

While I knew this would be the case intellectually, it still is rather stark to see the once might 780 get less than 20FPS in many modern titles.
 
MavericK said:
Could be other factors at work, but I would guess that much of that is due to the 1080 having double the VRAM of the 980, versus the 980 only having 1 GB more than the 780.
Click to expand...
The larger jump from the 980 to the 1080 is largely attributable due to the huge process node jump (780 = 980 = 28nm, 1080 = 16nm), combined with the fact that the 780 is a cut down 780 Ti (big chip) and not the fully enabled mid-sized chip like the 980 and 1080 (which the 770 would correspond to). The fact that the 980 is significantly faster than the 780 at all while still being 28nm is related to the clockspeed advantage on refined 28nm but most significantly due to the huge efficiency jump of the Maxwell microarchitecture (which is the main reason AMD is behind to the degree it is - Maxwell really was a kind of "quantum leap", similar to Conroe or Sandy Bridge in desktop CPUs).
 
In regards to the jump from Maxwell to Pascal.

It used to be a bit of a joke with Maxwell, just add +300mhz to the core and you would likely be able to stabilize that clock. With Pascal nVidia really pushed their TDP management to the next level, so you basically get that "extra" margin now on a "stock" card with the boost modes.

Just 2 cents in there.

That said, in my experience even with a +250mhz core clock boost on the 980Ti I had, it could "only" equal my 1070 (non-Ti). The 1070 would bench higher, but the 980Ti would keep a higher minimum framerate. I just give these as comparison points, I also could have been cpu bound.

Just some thoughts on Pascal vs Maxwell.
 
I refer to my purchase of a gtx780 as the worst PC component purchase I've ever made. As you point out, it was eclipsed in short order by games as well as Nvidia, seems to have had poor driver optimization support, and it cost more than the next 2 generation's cards.
 
enjoyed the Part 1 ... thanks.

Also noticed this:

"as we anxiously await what NVIDIA has next up its sleeve for 2018"

hmmmm ...
 
Last edited:
Well gawd daayaaaaaamn, I never would have expected to see such a massive jump in performance going from Maxwell v2 to Pascal, than going from Kepler to Maxwell v2. Geeeezus. The drop in size to the smaller node and the refinements and improvements made to the Maxwell architecture that turned it into Pascal really counted for a whole helluva lot, didn't it. As always thanks to tha [H] for the work it took to present us with this info. Looking forward to reading the next parts.
 
Love these kinds of articles, can't wait for the all in one comparison when you get done.
 
Cool comparison! Here's hoping the 1180 adds another 50%. I went form a 760 to a 1080 and at 4K plenty of games push the 1080 to the limit.
 
Thanks Brent&Kyle for all your hard work on this. I enjoyed as I've lived it. I also really liked how you used 1440p as the standard resolution. For the time being, 4k, isn't a realistic target for most. I'm pretty sure Part 2 will show some interesting results for 1440p. Got a giggle out of you using KCD for benching, this game makes me grumpy with the tricks I have to do to get to a playable 4k setting. Totally reminds me of Witcher 2 Uber at 1080p back in the day. So many settings, such a PIA to get playable. Nowadays just about any x80 will do the trick.

The rigs in my profile have had many GPU's. Most of which are in this article. Had 780>G1 970 SLI w/ 780 physx>G1 1080 SLI. The 780 I had was an EVGA SC780. Such a beast for the time and it had replaced a pair of PNY560TI's in SLI. Loved it. Only regret was not remembering about the TI's coming down the road. Was really happy with my G1 970's in SLI too, thanks to the SLI support of the time they had no trouble trouncing a single 980 and even kept up with a 980TI for most things in 2016. Never had a 980 but I do have a MSI Titan with OC'd 980m SLI. Sad part is they're roughly the same performance as the desktop 970's and 8GB Vram. Ironically the Vram size is pretty much useless at 1080p which is mostly much the limit of what the clocks can deal with in ultra for most games. The G1 1080SLI I have now still out paces most things in 4k compared to my single Strix 1080TI but Vram, as you noted with the 780, has continued to be an issue. MEA and ROTR both will hit the ceiling for them with AA type settings maxed, while the 1080TI has 'just enough' to get by but lacks the raw processing power to hold 4k/60hz.

From the 780's to 980's I used to say that if they'd just make it so you could upgrade the Vram then most of those gen's would last years more. Obviously defeating any point of anyone upgrading to a new card. No surprise either that SLI support fell off a cliff after the 970's since so many we're happy with even the 3.5 ceiling for most things. The performance/price ratio was amazing then.

I really want to jump ship to AMD in the next round but honestly whoever makes something a bit better than a 1080TI will get my money as the probable last card I put in my 4930k rig. For the next round I just want a single, 4k/60hz, above 11GB, solution. Assuming that MOBO survives, I'll probably upgrade to PCIe NVMe's at that point since I'll have the slots/lanes to spare. The only thing that will hold me back is $$$. If the rumors are true about being ~$1200-1500 then I'll stop here. It's just not worth that much to me anymore. A thousand is my limit.
 
I think this also explains why nVidia doesn't need to push the next gen card out so fast...they overleaped the step from the 980 and gave a little "too much". If the 1180 has as big a delta from the 1080 as the 1080 has from the 980...I would probably buy now and realize this is good as it gets for a while.
 
rumax said:
No GPU out there runs at 2 GHz out of the box right now, so I don't know what you mean about being "stuck" with 2 GHz. I also don't understand the obsession with figuring out IPC on a GPU when it isn't the most important metric of its performance in the least.
Click to expand...

Do some searching here at [H]ard and you'll find Kyle has given great insight on how to make this happen. My air cooled rigs, 1080TI, and 1080SLI both are running at over 2Ghz stable, 50-60c. It's really not that difficult, right amount of power, keep 'em under 60c, and honestly not usually a problem. Even the 1080SLI will hold 2012MHZ upwards of 65c. I only run less when I want the rooms as close to silence as possible.
 
To be even more accurate about the 2Ghz limit we could throw in liquid. I've seen some 'order-able' solutions, non ln, that can keep things around 40-50c but even then the ceiling for most is ~2.1-2.2Ghz. Current Pascal just can't seem to go past 2.2 stable but 2.0 is pretty easy.

edit: Again I have to say the Kyle/Brent have done extensive reviews on these solutions as well. The 2.0 limit is really just a generalization as to what many cards will do with minor tweaking 'out of the box'.
 
lostin3d said:
Do some searching here at [H]ard and you'll find Kyle has given great insight on how to make this happen. My air cooled rigs, 1080TI, and 1080SLI both are running at over 2Ghz stable, 50-60c. It's really not that difficult, right amount of power, keep 'em under 60c, and honestly not usually a problem. Even the 1080SLI will hold 2012MHZ upwards of 65c. I only run less when I want the rooms as close to silence as possible.
Click to expand...

Check's own signature, damn I must be lying.
 
rumax said:
No GPU out there runs at 2 GHz out of the box right now, so I don't know what you mean about being "stuck" with 2 GHz. I also don't understand the obsession with figuring out IPC on a GPU when it isn't the most important metric of its performance in the least.
Click to expand...

Heh. My GPU has been running at just over 2Ghz since August 2016.
 
You must log in or register to reply here.
Back
Top