Kepler sure did take a dive in this title -- that's an eye opener. Thanks for the link.
Even if we ignore AMD's and Kepler's performance though, check out that chart:
970 = 46 fps
980 = 47 fps
Titan X = 48 fps
The game's engine seems to have major CPU (other?) bottlenecks. This type of scaling is simply unacceptable. It's possible there are certain parts of this game engine going back to NFS Shift series that are hogging down performance on all GPUs.
Also, if you look at AMD's or NV's performance from 1080P to 4K, the performance drop off is too small to make logical sense (4x the pixels but performance often drops off just 15-31% for NV cards).
1080P, rain
Titan X = 48 fps
GTX980 = 47 fps
780Ti = 39 fps
4K, rain
Titan X = 41 fps (-15% drop)
GTX980 = 33 fps (-30% drop)
780Ti = 27 fps (-31% drop)
I am not a programmer but it appears there are some major CPU bottlenecks or game engine issues at lower resolutions in this game. There is no way going from 1080P to 4K should only be a 15-31% drop in FPS. To me it doesn't appear that this game is well optimized to start with.
I would like to see benchmarks of this titles after various developer patches, NV/AMD driver updates and all done on Windows 10 where CPU overhead is reduced. I bet performance will improve for everyone.
I believe there is a lot more to it than what is being repeated on the forums. Kepler performance has been hurting in a lot of recent titles, that is for sure.
Of course AMDs GCN isn't showing its age, it is still their current line up. It is AMDs current gen. On to of that, GCN is now in the current gen consoles. That was supposed to give them an edge and I think people just ignore that. Then, lets not forget that AMD had some pretty notable big driver improvements over the years.
Those are just a few things that any level headed person could easily see. Then there is the flip side. If people comparing recent performance of Hawaii and Tahiti to Kepler just put it off as nvidia abandoning Kepler, they the doing a disservice to AMD. AMD put a lot of effort in getting the most out of the gpus and it is finally paying off well.
I think Tahiti was a very beefy design. It had some killer specs. So much so that on paper there is no way the gk104 should have been able to touch it. It makes so much more sense to me that AMD just got better at harnessing that power.
I remember I did a mathematical calculation and I compared Kepler to Maxwell, Kepler to GCN and GCN to Maxwell and the result was that GCN narrowed its gap to Maxwell by a couple percent but in the same time Kepler fell far behind both GCN and Maxwell in the same period. It's hard to believe that almost immediately after GTX970/980 launched that by pure chance Kepler started bombing and especially so in GW titles.
For example, Computerbase has 960 just 6% within a 780Ti in their testing. NV has publicly stated that Maxwell has about a 35-40% increase in IPC per CUDA core (128 Maxwell cores are 90% as fast as 192 Kepler cores). It's impossible to explain how a 1024 CUDA core 960 can be so close in performance to 2880 CUDA core 780Ti unless NV improved IPC by
2.5X. There are so many titles now where the 960 is near 680's level of performance, 970 is beating 780Ti, 980 is far ahead of 780TI sometimes by 25-35%, 780 barely beating 7970Ghz!
Tessellation can't be the reason because Maxwell hardly improved it over Kepler. Memory bandwidth can't be the reason because 780Ti ha at least as much as a 970 even if we account for 33% more efficient use of memory bandwidth by the 970. Just look at the number of texture mapping units of 780Ti vs. 970. Unless all Kepler cards are suddenly pixel fill-rate limited, what is the explanation that driver focus by NV is now on Maxwell? Also, if pixel fill-rate was an issue, 7970Ghz would start taking even more but its lead over 680/770 keeps growing and its performance is barely behind a 780.
Some have said most of modern titles utilize a lot more compute shaders in the pipeline but I have yet to see any developer claim they are using a lot of DirectCompute in their games, especially not in GW games because that would instantly give GCN a major advantage even over Maxwell.
We probably won't ever know if NV stopped optimizing drivers for Kepler and refocused all its energy on Maxwell but for a consumer who bought $650 780, $1K Titan, $700 780Ti, it doesn't particular matter what the reason is -- when they launch the game and see a $180 960 nearly as fast as their $700 780Ti or the ancient 7970Ghz being a hair behind 780 and R9 290X now matching or beating 780Ti, I doubt they are happy about this. I don't recall GTX480/580 falling apart this badly just 1.5 years from their launch. GTX580 performed about as well as a 660Ti/7870 and started to suffer a lot later on due to 1.5GB VRAM.
That's all fair enough, ocre, and it'd be hard to tell some of that.
However there should probably be a thread comparing:
- GCN and Kepler once both had been launched
- GCN, Kepler and Maxwell at Maxwell launch
- GCN, Kepler and Maxwell now
All on consistent games released before that point. So 1 gets a few games, then 2 adds a few more and gives a second look at later performance in game set 1 and so on.
It would definitely be interesting to see how they fared next to each other, and whether for example, Kepler has gotten left behind without improvements or Maxwell has improved significantly more than Kepler.
Someone did a very extensive analysis of this and it does show Kepler performance fell off a cliff following 970/980's launch. For 12 months leading up to 970/980 launch, 780/780Ti performed well against 7970Ghz/R9 290/290X.
https://www.reddit.com/r/buildapc/comments/2one2z/discussion_has_nvidia_forsaken_kepler_cards_has/
"General conclusions
First of all, the relative performance of Nvidia cards compared to AMD cards, for both Maxwell and Kepler, is down across the board. For example, at 1080p, the 970 which previously had an 8.0% lead over the 290X now finds itself only 4.1% ahead. The Titan, which had a 2.3% performance advantage over the 290X, now finds itself trailing the AMD card by 4.1%. At 4K, the story is similar: the 980’s 11.2% performance lead over the 290X is down to 8.6% and the 780 Ti, which used to beat the 290 X by 3.1%, is now losing by 1%. However, another conclusion that can be drawn is that the relative performance of Kepler cards has dropped by much more than their Maxwell counterparts’. At 1080p, while the 980 goes from beating the 290X by 22.7% to having a 17.3% lead, the 780 Ti’s 14.7% lead over the 290X drops to a lowly 6.1%. This can be seen by comparing the two Nvidia cards, and the 980’s 6.9% performance advantage over the 780 Ti grows to 10.6%. And at 4K, while the 970’s deficit over the 290X grows from 3.1% to 4.8%, the Titan’s 5.1% deficit turns into a huge 11.4% loss to AMD’s card. The 970’s relative performance thus grows from beating the Titan by 3.3% to 7.5%. These numbers might not seem significant, but bear in mind that they are caused by a handful of new games."
Take a look at Sweclockers
Sept 19, 2014 - Maxwell launch
980 is 16% faster than a 290X (Uber)
780Ti is 9% faster than a 290X (Uber)
Titan is 29% faster than 280X
vs.
980 is 11% faster than 290X (Uber)
780Ti is tied with 290X (Uber)
Titan is 20% faster than a 280X
Kepler lost way more performance than Maxwell vs. GCN since Sept 2014. Whether Kepler architecture is becoming outdated, or NV is not focusing on the drivers as much -- someone who bought last gen's $600+ Kepler cards can't be happy with this!
That's why I keep advocating for PC gamers to upgrade more often than trying to buy $600-700 cards to future-proof for 3-4 years. You never know if AMD's or NV's architecture will get crippled by drivers/bottlenecks in 2 years. Historically though, if we go back to HD4890 vs. GTX280 and until today's line-ups, AMD cards tend to perform better in more demanding next gen titles as they age, while NV's advantage more or less disappears from 15-20% to single digits (think 6-7%) as time goes on. The one big exception to this rule is the HD5850/5870 series which got killed against 470/480/580 cards due to poor tessellation and having only 1GB of VRAM.