The GTX 780 TI & SLI - End of Life - Performance Analysis
The point of this post / thread was to display some hard facts and anecdotal experiences with the GTX 780 Ti SLI.
Kepler Degradation In Performance:
For something that could be tested, it seems like a lot of people have not taken the time, an arduous and tedious amount of time, to produce some tangible data regarding Kepler degradation in performance. Popular opinion even states that the last viable driver was 347.88. Well I am here to tell that I have data that may prove otherwise.
I am going to let you guys make your own conclusion, I have no interest in persuading opinion towards nVidia or even against nVidia. There are somethings that I feel should be corrected, because when people have all the facts, they can make informed opinions themselves.
My motivation is just help inform users and I feel obliged to do the testing since I own the two cards.
The list of drivers I have tested:
3DMark Fire Strike
3DMark Fire Strike Extreme
Why the one benchmark? Well initially I had planned to do a few benchmarks. The difficulty arises in my mind when trying to find a benchmark that's been around a while and supported in SLI by multiple drivers. For example GTAV is only good for 3 previous drivers and that wouldn't show enough data in my mind. 3DMark is also something that sells GPUs. This is the de facto standard for reviewing cards/benchmarking cards. In my mind, this is where nVidia, would at least focus support every driver release, and has to be at least one of the in-house tests they use to gauge performance of a driver. This led to the following idea...
nVidia has to dedicate time and support to testing 3DMark every release. Any substantial changes in performance should cascade outwards, especially negative performance.
If 3DMark drops measurably in performance. This would confirm larger, measurable, performance issues in titles with less focus could exist.
If 3DMark does not change in performance, other titles are not guaranteed to be reflective of this. More testing is needed then.
So I currently only have used one benchmark, what do you think that means ? This article may or may not be finished, I may go back further and testing more drivers.
Three runs per test.
Six runs per driver.
Scores below are averaged of three. GPU scores recorded and compared separately.
Blue Scores - Fire Strike Scores
Green Scores - Fire Strike GPU Scores
Results SLI (Charts):
Results SLI (Numbers):
353.38 – A huge improvement over 353.30 in terms of stability and performance.
353.30 - Just could not get a 3rd run on 3DMark Fire Strike Extreme tried 5 times even on stock w/ reboots. GPUs do not idle (go to a lower power state) once overclocked.
353.06 - GPUs do not appear idle (go to a lower power state) once overclocked.
352.86 - GPUs do not appear idle (go to a lower power state) once overclocked.
350.12 - GPUs do not appear idle (go to a lower power state) once overclocked.
347.88 - GPUs do not appear idle (go to a lower power state) once overclocked.
347.52 - This is driver does not have the 20 seconds freeze in NVCP when changing 3D settings. GPUs idle (enter a lower powerstate) even with OC set!
347.25 - 3DMark won't launch if Afterburner OSD is running. Changes to the NVCP 3D Settings don't seem to be sticking i.e. Prefer maximum performance, however scores are the highest.
Overclocking Retested on the Latest Drivers (353.30):
001 - Driver crash & recovered in BF Hardline before testing. Driver recovered successfully but obviously stuck in the wrong power state.
002 - Didn't even make past the loading screen. No overclock, just a DirectX crash. PowerLimit seems to trigger it. Reset PowerLimit to default.
003 - http://www.3dmark.com/3dm/7484258?
004 - http://www.3dmark.com/3dm/7484455?
005 - http://www.3dmark.com/3dm/7484578?
006 - http://www.3dmark.com/3dm/7484642?
007 - http://www.3dmark.com/3dm/7484712?
008 - http://www.3dmark.com/3dm/7484801?
009 - Driver crash. Conducted a reboot before next test.
010 - http://www.3dmark.com/3dm/7484960?
011 - Driver crash. Realized voltages are not increasing in Afterburner. Set Unlock Voltage to Kernel Mode, Standard MSI. Working again.
012 - Switched PrecisionX. Driver Crash.
013 - http://www.3dmark.com/3dm/7485714?
014 - http://www.3dmark.com/3dm/7485797?
015 - http://www.3dmark.com/3dm/7485872?
016 - http://www.3dmark.com/3dm/7486063?
017 - http://www.3dmark.com/3dm/7486150?
018 - http://www.3dmark.com/3dm/7486237?
019 - http://www.3dmark.com/3dm/7486328?
Why is the 780 Ti struggling against a card it used to beat?
A.) Well, we are EOL, I would assume that the 3 years of Kepler optimizations have already been made.
I expect new title support, but I don't expect the same level of focus on my GPU anymore.
B.) Maxwell architecture is new.
Consistent newer title reviews have shown the GTX 980 / 970 are currently ahead of the GTX 780 Ti. Newer titles require newer drivers.
Most claims that the GTX 780 Ti used to beat the GTX 980 / 970 are made based on launch reviews. Older drivers were used.
New architectures mature with driver release.
Benchmarks and titles where Tessellation are heavily taxed, Maxwell has 3x tessellation performance as Kepler.
The GTX 970 and 980 also have another hardware advantage, another 1GB of VRAM, or 512 MB of VRAM in the 970 case.
C.) GCN architecture isn't new!?
AMD has one architecture to support. Optimizations for the biggest GCN card trickle all the way down to the smallest GCN card, or even older cards.
Most claims that the GTX 780 Ti used to beat the R9 290X are based on old launch reviews. Older drivers were used. Newer titles require newer drivers.
All architectures mature with driver releases. Even GCN will be tapped out at some point in the future.
D.) The transition to the 35x.xx family has left us a little down in terms of performance.
Windows 10 / DX12 driver changes are slowly being merged into 34x.xx family of drivers. Bugs will come up first, optimizations always come afterwards, if they come at all.
They could have done a better job in merging the branch changes in the drivers.
E.) All The Console Ports Have Been Virtually Terrible
Batman: Arkham Knight – Sales have been stopped it was so bad.
Assassin’s Creed Unity – Appalling and terrifying release.
Watch_Dogs – Hatefully now known as Patch_Dogs.
VRAM requirements have tripled over night with no perceivable increase in image quality.
The architecture in the consoles GPUs is GCN. Games are being designed from the ground up to work well on AMD hardware.
Newer games simply will have a few more low level AMD optimizations at launch than nVidia ones.
Conclusion (as of 7/3/2015):
The transitioning into 35x.xx family has caused a certain amount of performance penalty to some Kepler users, for example a loss of 4.05% in GPU score in Fire Strike, and 3.88% Fire Strike Extreme. A loss though should be read in context. The loss in performance since at least driver 347.25. I don't think there is a huge conspiracy to "withhold" performance as people speculate, but there is enough data to merit a review by nVidia. I don't expect major performance gains any more, now that we are EOL, but we shouldn't be going down in performance in a benchmark that even pre-dates this video card.
The best driver currently for Kepler, in terms of 3DMark performance I have tested thus far is 347.25. If you must use a 35x.xx driver, I highly recommend the 350.12 or the 352.86. I have updated the charts and added the actual numbers.
Updated to Conclusion:
The numbers since 340.52 only show a lost of 3.96% loss of performance in the GPU score, or 2.09% in overall score in Fire Strike. The latest hotfix driver 353.38 seems to be that nVidia is taking the issues serious and managed to resolve not only the performance issues but the vast majority of TDR crashing that was occurring.
Also note: The 353.30 is the worst driver I have ever used from nVidia and I have had Quad SLI and GTX 590s before.
I have greatly enjoyed my time with the GTX 780 Tis, even though they are not good overclockers. Out of all my nVidia experiences, they have been my favorite cards. nVidia driver stability has been usually good for me, that being said, the 353.30 is the worst driver I have ever used from nVidia and I have had Quad SLI and GTX 590s before.
I have been using the 780 Tis since November 2013. I did my testing in my configuration because this has been my experience with Kepler GPUs in SLI. To me there hasn’t been a mass reduction in performance, I have never felt this crushing ominous crippling of performance that so many people are claiming is happening. I don’t doubt there are issues, issues which seem to vary based on each user’s configuration (including driver issues) but as far as I can tell no one has taken the time to prove many of these claims. I recommend everyone, if they can, to use the 347.25 driver, it is a solid performer and very stable. Maxwell also predates this driver which refutes the whole “nerfing Kepler to make Maxwell shine” theory. Looking at the trend up to the 347.25 driver, performance has been consistent. I personally do not have the TDR bug in Windows 8.1, I did have it in Windows 10, it does look like they are working on a fix finally though.
The tests here are for a narrow premise and not all encompassing, I realize that. Again the point of this article is encourage neutral fact based discussion but also provide a reference to repeatable and reproducible tests. All my test results are under the user RagingCain at 3DMark Results, if you want/need that extra proof.
My Driver Recommendations In Order
347.25 – nVidia control panel settings do change, but when the page refreshes only the default choices appear to be selected.
340.52 (For Non-SLI) – Performs slightly better in single GPU setups than 347.25, but not SLI.
353.38 – Have not tested Shadowplay or Geforce Experience. No freezing loading the nVidia Control panel, settings do stick.
350.12 – Had difficulties using Shadowplay.
Article Source: ByteMeDev.com
Coming Up Next
Grand Theft Auto V when they fix the latest performance issues.