[PcGameshardware] The Witcher 3 Benchmark

[RussianSensation]

The TechPowerUp article posted above explains the Kepler issue as related to tesselation. I wonder if that's the only cause.

That's the most unfounded explanation and I am surprised W1zzard who reviews GPU hardware didn't know that 780/780Ti have very strong geometry performance.

770 > 960 at tessellation. This can't be the answer unless the driver is at fault.

The focus on Maxwell was not on geometry performance and 780Ti isn't far behind the Titan X.

Even if we increase tessellation to 64X, 780Ti clobbers the 290X and isn't far behind the 980/Titan X. That means not only is tessellation the wrong answer, it can't be the reason unless the driver is fully unoptimized for Kepler. 980 and 780Ti have nearly the same tessellation performance at 64x.

770 > 960 at texture fill-rate.

780Ti > 980 at texture fill-rate. This also can't be the answer.

Pixel fill-rate is the shared weakest link in Kepler and GCN 1.0/1.1 compared to Maxwell. This might be one reason for Kepler but it hardly explains why R9 290/290X > 285 despite 285 beating them in pixel fill-rate.

R9 280X is better at compute than R9 285 as well.

http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/17

It's really hard to figure out right now what's causing the performance to tank so much on GCN 1.0 and Kepler in TW3.

 

[96Firebird]

I still believe Kepler is falling behind because of Forward+ rendering and it's lack of DC power...

We see the 780 Ti below the 290 in TW3, which uses Forward+ rendering techniques. We see it on par with a 280X in Ryse, another game that uses Forward+ rendering. Again below a 290 in Evolve, which also uses Forward+. DiRT Rally I can't explain, not sure if resources are freed up because the game is not very GPU demanding?

Looking at other recent games, GTA V has the 780 Ti on par with a 290X. In BF Hardline it is above the 290X @ 1080p, on par @ 1600p. Dying Light it is above the 290X.

Looking at TPU performance summary for the Titan X review, the 780 Ti trades blows with the 290X.

I got all my figures from looking at GameGPU tests, FWIW.

 

[RussianSensation]

I still believe Kepler is falling behind because of Forward+ rendering and it's lack of DC power...

Where is the evidence that 780Ti is way worse in direct compute vs. 980, especially when it comes to Forward+?

http://www.computerbase.de/2014-09/...-nvidia/9/#diagramm-gpu-computing-computemark

Also, that's a BOLD claim you are making considering 960 is just 6% within a 780Ti in Project CARS - that forward+ is the most important reason why Kepler tanks. 960 is worse than a 780Ti in everything. You are also ignoring many other games where Kepler cards started tanking as early as last year. It's not just limited to Forward+ cases.

Your theory actually doesn't work at all because if compute was a factor, and you site Evolve and Ryse Son of Rome as 2 examples of how Forward+ cripples Kepler, then how do you explain this?

780Ti and 970 are more or less tied in Ryse Son of Rome.

780Ti > 970 in Evolve.

You must have used outdated GameGPU data with older drivers for both of those games and without the games re-tested with after they were patched. That's one of the major downfalls to a lot of GameGPU testing - they rarely retest games with newer drivers/after the game has been patched 5-6 times. That's why we need to look at more than 1 site for updated testing with newer drivers, etc.

Other situations can't at all be explained by DirectCompute either. How can 960 be 6-10% within 780Ti in Project CARS but R9 290/290X have such poor performance? Sorry but Compute cannot explain that since 780TI has 2880 stream processors and R9 290/290X crush 960 in compute. Therefore, compute again can't at all explain Maxwell vs. Kepler performance in Project CARS and at the same time reconcile the poor performance of R9 290/290X that excel at Compute.

Also, 970 and 780Ti are trading blows in Dragon Age Inquisition, another modern game but yet R9 290/290X are faster than a 970 in DAI but 970 beats 290/290X other GW titles that you suggest are "heavy" on compute? What makes you think 1 game is Compute heavy and another isn't?

There are too many examples where 780/780TI started performing worse well before TW3 and in games that don't have any evidence of being DirectCompute heavy.

Another reason your compute argument doesn't explain it is you ignore the AMD side. 280X is superior in Compute to a 285 but 285 wins in TW3 for example but loses in just about every game to a 280X besides Bioshock Infinite.

 

[Hi-Fi Man]

This reminds me of when G80 and it's derivatives were released. G70/G71 was a good uArch but as time rolled on and games started to take advantage of the unified arch and larger back end G70/G71 looked very dated.

The GeForce 7900 GTX was faster than the 8600 GTS at first however, the 8600 GTS ended up being faster.

The situation now isn't quite the same or as drastic as it was back then however, it's obvious Gameworks caters towards Maxwell. I believe this is in part due to uArch differences and focus of optimizations for Maxwell first then Kepler later.

As RussianSensation pointed out it's not tessellation or directcompute performance. I think as far as uArch goes it's to do with single precision FP performance and pixel fill rate. Maxwell's single precision FP performance and pixel fill rate is superior to Kepler's according to Anand bench data.

It seems pixel fill rate is becoming far more important again (like it was in the 90s and early 00s). I think some people need to realize that goalposts are always changing for the GPU makers and this is reflected in their uArch designs.

It also seems some people aren't happy that their two SLI 780 Tis that they spent big $$$ on aren't performing well with GameWorks enabled. Honestly they should be happy that these cards still perform well without Hairworks enabled. 10+ years ago graphics cards were truly being made irrelevant (not being able to play the latest games at all) in a matter of a year or two. An example would be the Radeon X850 XT, this card was a pixel shader 2.0 card. Shortly after this card came out the shader model 3.0 revolution happened (around the time the Xbox 360 became popular). This rendered the card almost completely obsolete in a matter of a year or two. The Radeon X850 XT was considered a good card at the time and it wasn't chastised because people back then knew that shader model 3.0 was the future and that R400 was going to obsolete. Things were even worse during the late 90s.

Now I'm not saying that Nvidia shouldn't be scrutinized (they should) for keeping Kepler in the dark in regards to the latest GameWorks features however, what I'm saying is that people should understand that things change and architectural deficiencies become exposed over time and that maybe next time you decide to spend big $$$ on a top card you should think a little bit more. It's for this sole reason I will never spend over $300 on a card.

On a side note, the mud slinging needs to end. Things should be discussed and debated without resorting to childish behaviour. This is a technical forum after all.

 

[RussianSensation]

I ran some benchmarks on three of my cards, a GTX 980, GTX 780 Ti, and an R9 290, and came up with some pretty surprising findings. Here's my blogpost about it.

In short, it's not Gameworks that's causing the problem for Kepler. But Kepler is definitely under-performing in this game, so something in the game engine is hitting it really hard.

Thoughts?

No idea really. 980 is 52% faster than a 780Ti while 290 ties a 780Ti in your testing. Minimum frame rates on your 780Ti are even worse than a 290, a card that was $400 vs. $700 for the 780Ti. Ridiculous.

It's quite possible Kepler and GCN 1.0 are starting to run into pixel fill-rate bottlenecks.

While 3DMark pixel fill-rate test shows 780Ti hangs in there with the 290X, it depends on the how optimized the program is.

But it's not that simple it seems. In other cases, Kepler's pixel fill-rate performance bombs which sounds like Kepler needs a lot of driver optimization/application to work well with it to extract maximum pixel fill-rate. Otherwise, we get this:

^ If NV stopped optimizing drivers for Kepler, all of a sudden it's possible (I am not saying 100%) that the pixel fill-rate bottleneck is exposed for all Kepler products. That somewhat explains how R9 285 can beat 280X in TW3 and why Kepler bombs in the same game against R9 290X or Kepler's 980. Essentially if the card is pixel fill-rate /ROP limited, its memory bandwidth, tessellation, shader performance advantages are all wasted since they are underutilized. If pixel fill-rate limitation comes into play, that covers why both the 280 and Kepler are starting to fall apart in modern games even without tessellation penalties. We know tessellation is a non-factor for Kepler since Kepler's is nearly as good as Maxwell's! The pixel fill-rate bottleneck would also explain why 780Ti hardly scales above 780 in some modern titles.

I still think it's more complicated. However, as a gamer/consumer, we don't care what the reason is. Fact is $650 and $700 780/780Ti fell apart in 1.5-2 years. If anything this reinforces the idea that trying to future proof with $700-1400 cards > 2 years is a waste of time and $. Updating more often to the latest architectures is the best way to future-proof since we can be sure AMD/NV will support/focus on those new products.

The answer could also be as easy as driver optimizations. In the Witcher 3, performance is all over the place for Kepler and GCN cards. 780Ti can trade blows with a 970 but at other times it loses to a 290. But also 290X goes from beating a 970 to losing to it badly. Also, 980 is 25-26% faster than a 970? Seems BS or some odd benchmarks.

This site always has the most strange results.

I haven't found a single site with The Witcher 3 benchmarks that seems to represent the game well enough since it seems performance varies greatly depending on the area/section of the game.

 

[Silverforce11]

^ If NV stopped optimizing drivers for Kepler, all of a sudden it's possible (I am not saying 100%) that the pixel fill-rate bottleneck is exposed for all Kepler products.

I still think it's more complicated. However, as a gamer/consumer, we don't care what the reason is. Fact is $650 and $700 780/780Ti fell apart in 1.5-2 years. If anything this reinforces the idea that trying to future proof with $700-1400 cards > 2 years is a waste of time and $. Updating more often to the latest architectures is the best way to future-proof since we can be sure AMD/NV will support/focus on those new products.

Because game performance is so dependent on drivers, if optimization is halted, performance will tank.

Also, someone who bought R290/X on release, its still going strong now and likely for another 2 years. So your theory doesn't apply to AMD GPUs, because GCN will still receive attention as its the underlying architecture for all of AMD's APU/GPUs & consoles for awhile yet. For example, given the poor performance of Kepler since Maxwell arrived, I would be more assured that my money buys more future proofing in a 390X than a 980Ti/Titan X for example, because once Pascal arrives, what happens to Maxwell, will it receive the Kepler non-treatment?

 

[Termie]

...

^ If NV stopped optimizing drivers for Kepler, all of a sudden it's possible (I am not saying 100%) that the pixel fill-rate bottleneck is exposed for all Kepler products. That somewhat explains how R9 285 can beat 280X in TW3 and why Kepler bombs in the same game against R9 290X or Kepler's 980. Essentially if the card is pixel fill-rate /ROP limited, its memory bandwidth, tessellation, shader performance advantages are all wasted since they are underutilized. If pixel fill-rate limitation comes into play, that covers why both the 280 and Kepler are starting to fall apart in modern games even without tessellation penalties. We know tessellation is a non-factor for Kepler since Kepler's is nearly as good as Maxwell's! The pixel fill-rate bottleneck would also explain why 780Ti hardly scales above 780 in some modern titles.
....

Interesting theories, RS. You definitely bring up some good points.

I'll just piggyback on what you said and mention something odd I noticed in my benching of the 780 Ti in Witcher 3: while GPU usage was pegged at 99% the entire time, the card wasn't operating anywhere near its power limit. I've never seen this behavior before, as 99% usage typically forces the card right against the power limit, leading to throttling in some cases.

I can't explain this other than to say whatever component of the 780 Ti was being pushed to its limits was not the most power-hungry component. I think that rules out tessellation as a factor here, based on my experience with power use during tessellation-heavy scenes.

 

[hawtdawg]

been playing with a gamepad on my tv. everything ultra with Lagworks off and AA off, and i have never seen this game dip below 60fps on my 780's

 

[96Firebird]

Where is the evidence that 780Ti is way worse in direct compute vs. 980, especially when it comes to Forward+?

I never said anything about the 980, I am comparing the 780 Ti and the 290 and 290X.

Evidence comes from our very own Ryan Smith at Anandtech, who writes articles in English that I can understand!

DirectCompute is the compute backend for C++ AMP on Windows, so this forms our other DirectCompute test.

Source

I was going to respond to your post before you edited into a wall-of-text, just pointing out a theory and supplying some evidence. You seem very upset by it, however, best not to take it personally...

Edit - I'll be civil...

 

[gamervivek]

980 does seem to do about 20% better than 970, quite a difference between the cards.

 

[Silverforce11]

Compute != Compute. Just becareful because Bitmining is also a GPU Compute scenario.

It depends on what the task is.

@96Firebird
I didn't get the "upset" tone of RS's post, it was very informative and backed up by valid proof.

 

[RussianSensation]

ISource

I was going to respond to your post before you edited into a wall-of-text, just pointing out a theory and supplying some evidence. You seem very upset by it, however, best not to take it personally...

Edit - I'll be civil...

I wasn't upset at all. As I was looking up data, I was editing my post and adding more examples. I don't know what the reason is but I don't believe compute is that reason. You site C++ as an example of DirectCompute where 980 is only 20% faster than 780Ti.

GameGPU added new data where they tested cards without Hairworks and no HBAO+ (ironically both features AMD's current drivers can't optimize for).

And what do we get?

290X > 970
290 > 780Ti
290 > Titan
280X > 780

At 4K, 780Ti is just 1 fps faster than a 280X and 280X beats the Titan. I don't think compute explains this.

This game does not even use 3GB of VRAM so 780Ti's is not running into a VRAM bottleneck.

*** Ironically The Witcher 3 is a game that 100% proves that GameWorks features unfairly penalize AMD GPUs because GW obfuscates game code from AMD's driver and prevents the developer from optimizing the game code for AMD/Intel GPUs as well***

From what I've read, 8xMSAA is enabled by default when HairWorks is ON. I guess no one at NV thought about the trade-off of performance vs. IQ!

The minute Gameworks features are disabled in TW3 (HBAO+ and HairWorks), all GCN products perform as intended.

We can debate the reasons why Kepler performs poorly in modern GW titles, but the key trends remain - Kepler bombs in all GameWorks games I can think off relative to Maxwell. In vendor agnostic titles, Kepler performs well for the most part. We won't know what the real reason is but Kepler owners are in a world of trouble now.

If Kepler's performance dramatically improves in TW3 or other GW games with newer drivers, that means NV neglected Kepler all along and focused on Maxwell instead. This would undermine NV's reputation of having solid drivers. That means it's in NV's best interests to keep its mouth shot and maintain the positive PR that their software is not to blame. If Kepler driver optimization are the cause, NV would rather pretend drivers are not at fault because at least no one would be able to prove that NV's drivers were crippling Kepler all along. For long term reputation of NV, it's better to throw Kepler under the bus than to release performance increasing drivers and admit neglect all along. Side benefit is most Kepler owners will upgrade to NV because most gamers will just think that newer games made Kepler an outdated architecture.

I personally feel that gamers who paid $380-450 for 770 2-4GB, $500-650 for 780, $700+ for 780Ti or $1000 for OG Titan are stuck with outdated products and got ripped off. What the reason is doesn't matter really. Fact is a November 2013 780Ti is bombing in GW titles, while cards like 780 and Titan are barely beating a 280X. Ouch.

 

[96Firebird]

Sounds like you're looking for one single thing that explains the performance (or lack thereof) of Kepler, when it is probably multiple reasons that add up.

 

[Silverforce11]

For whatever reasons unknown to us because we don't have access to GameWorks development process. The fact that Kepler is doing very well outside of GameWorks titles show that there is something anti-Kepler with the GW program.

It could be compute, could be fill-rate etc, we just don't know why but we know the penalty is there.

ps. Sniper Elite & Sleeping Dogs is also Forward+ rendering IIRC. NV performs fine in those games post-driver updates to optimize for it.

 

[RussianSensation]

Sounds like you're looking for one single thing that explains the performance (or lack thereof) of Kepler, when it is probably multiple reasons that add up.

No, I am not even looking for any 1 reason. I am instead looking at the big picture -- Kepler products have poor performance in GW titles. I don't really care what the reason is because I am not a GPU designer/engineer. As a consumer, if I owned a Kepler product, I would be pissed right now. Also, as Silverforce pointed out, with newer drivers, 780/780Ti/OG Titan's performance in Forward+ games like Sniper Elite 3, Hitman Absolution, Sleeping Dogs, Dirt Showdown was all fixed.

That's why it's hilarious when people talk about GM200 6GB future-proofing and how NV's drivers are amazing. The marketing focus is Maxwell vs. R9 200 series, ya that series that was always meant to compete with GTX700 products, but outside of TechSpot, every single North American hardware site is keeping its mouth shut on how Kepler is bombing in games released in the last 8 months.

I think I would be in rage zone if I purchased dual $1000-1300 780s or $1400 780Tis last gen. Their performance today is atrocious.

780 SLI is just 4-5 fps faster at 1440P than 970/290X. Once AMD has a driver with CF support, it's going to turn into a situation where 280X CF is barely lower than 780Ti SLI.

 

[Serandur]

No, I am not even looking for any 1 reason. I am instead looking at the big picture -- Kepler products have poor performance in GW titles. I don't really care what the reason is because I am not a GPU designer/engineer. As a consumer, if I owned a Kepler product, I would be pissed right now. Also, as Silverforce pointed out, with newer drivers, 780/780Ti/OG Titan's performance in Forward+ games like Sniper Elite 3, Hitman Absolution, Sleeping Dogs, Dirt Showdown was all fixed.

That's why it's hilarious when people talk about GM200 6GB future-proofing and how NV's drivers are amazing. The marketing focus is Maxwell vs. R9 200 series, ya that series that was always meant to compete with GTX700 products, but outside of TechSpot, every single North American hardware site is keeping its mouth shut on how Kepler is bombing in games released in the last 8 months.

I think I would be in rage zone if I purchased dual $1000-1300 780s or $1400 780Tis last gen. Their performance today is atrocious.

780 SLI is just 4-5 fps faster at 1440P than 970/290X. Once AMD has a driver with CF support, it's going to turn into a situation where 280X CF is barely lower than 780Ti SLI.

Indeed. I'm still pretty mad even though I didn't keep my 780. What if I wanted to keep it? What if I want to keep my Maxwell card when Pascal releases, will its performance start falling apart unduly as well? I feel especially bad for my friends and family with Kepler cards or expensive Kepler laptops.

 

[VulgarDisplay]

Lol, so gtx780 SLI is on par with a single 290x.

How could anyone not come to the conclusion that nvidia is forcing upgrades on people by intentionally crippling their own cards? I'm glad I decided to return my gtx780 when it's fan crapped out after a week and got a r9 290 instead.

 

[RussianSensation]

For whatever reasons unknown to us because we don't have access to GameWorks development process.

Oh now the real picture is emerging. First GameGPU retested the game without Hairworks and it shows 290X > 970, 290 > 780Ti and 280X > 780.

Now Computerbase stepped up to the plate.

Legend for clock speeds:

R9 290X "OC" = 1030 mhz
R9 290 "OC" = 1007mhz
***In the past, Computerbase used reference 290 clocked at 862-901 MHz and 290X clocked at 838-871 MHz. :sneaky:

980 "OC" = 1,177 MHz base clock (reference card only has a base clock of 1126mhz)
970 "OC" = 1,184 MHz (reference card only has a base clock of 1050mhz)

Without HairWorks:

Titan X = 59.5 fps
980 = 49.1 fps
290X = 44.1 fps
290 = 41.2 fps
970 = 40 fps
780Ti = 35.4 fps
780 = 29.6 fps
285 = 28.2 fps
280X = 28.1 fps
http://www.computerbase.de/2015-05/...diagramm-grafikkarten-benchmarks-in-1920-1080

Essentially the game is ONLY playable at 1080P at good fps without Hairworks unless one is packing 970 SLI, 980 SLI or Titan X/Titan X SLI.

With Hairworks aka AMD Cripple-works:

Titan X = 49 fps
980 = 40.2 fps
970 = 33.7 fps
780Ti = 28.5 fps
290X = 28.2 fps
290 = 26.8 fps
780 = 24.1 fps
770 = 20.9 fps
280X = 20.3 fps
285 = 19.4 fps
http://www.computerbase.de/2015-05/...diagramm-grafikkarten-benchmarks-in-1920-1080

^^^ Who wants to play at those frame per second at 1080P?

 

[x3sphere]

Lol, so gtx780 SLI is on par with a single 290x.

How could anyone not come to the conclusion that nvidia is forcing upgrades on people by intentionally crippling their own cards? I'm glad I decided to return my gtx780 when it's fan crapped out after a week and got a r9 290 instead.

No, I don't think that's the case - I think there is a technical reason for it. But either way, it doesn't change the fact that Kepler buyers got burned. Myself, I am now skeptical Maxwell will retain its performance level when the shift to Pascal begins.

It will be interesting if NV responds as this has started to get a lot more attention on other forums with the release of TW3.

 

[VulgarDisplay]

No, I don't think that's the case - I think there is a technical reason for it. But either way, it doesn't change the fact that Kepler buyers got burned. Myself, I am now skeptical Maxwell will retain its performance level when the shift to Pascal begins.

It will be interesting if NV responds as this has started to get a lot more attention on other forums with the release of TW3.

There are some very big anti nvidia threads all over reddit PCMR lately. People are super upset about their kepler cards.

 

[RussianSensation]

Indeed. I'm still pretty mad even though I didn't keep my 780. What if I wanted to keep it? What if I want to keep my Maxwell card when Pascal releases, will its performance start falling apart unduly as well? I feel especially bad for my friends and family with Kepler cards or expensive Kepler laptops.

That's why I am surprised how many people are talking about keeping 390/GM200 cards for 2.5-3 years. Gamers should just buy those cards for today's games but not expect them to max out 2017-2018 PC games. If this was a forward looking new gen with brand new architecture and a node shrink, I'd understand. However, it's still the same GCN and Maxwell architecture, still on 28nm node. From my perspective, it's not about NV vs. AMD but how we as gamers should 'protect' ourselves from wasting $ so to speak. GPU makers want us to spend more $ to "future-proof". That's their marketing game. It's time we re-evaluate how we upgrade GPUs and maybe it's no longer a good idea to buy flagship card to keep them for 3-5 years. I never agreed with this strategy to be honest and I was always of the view that the best way to future proof is to upgrade sooner to a new architecture/node. What makes it a "double hit" for us if we either have to pay $400-500 for mid-range next gen products (ala 670/680/970/980/GP204) or upgrade every flagship gen. It's hard to say what the best way is anymore as they are pros and cons for each of them.

I am also surprised that tessellation is an all or nothing setting in games. It's been a long time since tessellation was debuted and we should have a setting of 2-64x or Low, Medium, High, Ultra but it seems tessellation in games today as a GW feature is either crazy overkill or none at all. Seems like an outdated concept to enforce a single tessellation factor, especially when the most extreme one is used. It doesn't seem like an effective way to manage performance vs. IQ options.

 

[VulgarDisplay]

That's why I am surprised how many people are talking about keeping 390/GM200 cards for 2.5-3 years.

I am also surprised that tessellation is an all or nothing setting in games. It's been a long time since tessellation was debuted and we should have a setting of 2-64x or Low, Medium, High, Ultra but it seems tessellation in games today as a GW feature is either crazy overkill or none at all. Seems like an outdated concept to enforce a single tessellation factor, especially when the most extreme one is used. It doesn't seem like an effective way to manage performance vs. IQ options.

You can force lower tesselation through drivers on AMD cards.

https://www.reddit.com/r/witcher/comments/36jpe9/how_to_run_hairworks_on_amd_cards_without/


Other tweak:


http://steamcommunity.com/app/292030/discussions/0/617335934150506844/#p3

Now I just wish someone would compare IQ with 8x, 16x, and in game Hairworks. Seeing that the jump from 8x to 16x is negligible, I doubt that there's a big jump in IQ to in game, and it's intentionally overdone to cripple older cards.

 

[chimaxi83]

Let's be honest. Hairworks in action looks so extremely meh, that I don't understand the performance hit. That, and why the hell is his hair weightless and permanently wavy? Wtf. Only thing benefiting from it is good ole Roach. They purposely made his poor non-HW mane terrible.

 

[RussianSensation]

You can force lower tesselation through drivers on AMD cards.

Now I just wish someone would compare IQ with 8x, 16x, and in game Hairworks. Seeing that the jump from 8x to 16x is negligible, I doubt that there's a big jump in IQ to in game, and it's intentionally overdone to cripple older cards.

Ya, I know that. AMD developed this feature in their drivers after Fermi smashed their HD5850/5870 cards in tessellation. My point is why can't game developers implement a varying slider for tessellation levels? It should be built into the settings menu in much the same way we adjust draw distance, texture quality, etc.

Let's be honest. Hairworks in action looks so extremely meh, that I don't understand the performance hit. That, and why the hell is his hair weightless and permanently wavy? Wtf. Only thing benefiting from it is good ole Roach. They purposely made his poor non-HW mane terrible.

Apparently NV sighted that as many as 30 enemies benefited from HairWorks.

 

[Abwx]

That's the most unfounded explanation and I am surprised W1zzard who reviews GPU hardware didn't know that 780/780Ti have very strong geometry performance.

770 > 960 at tessellation. This can't be the answer unless the driver is at fault.

The focus on Maxwell was not on geometry performance and 780Ti isn't far behind the Titan X.

Not related to the discussed issue but curious that in the first pic the 290 is at 17759 and jump suddenly to 27767 in the second pic, a score higher than the 290X in the third pic...

Not the same for the nVidia cards from first pic, the oldest, to the most recent one, quite indicative of Techreport usual shilling...