Is Nvidia deliberately downgrading Kepler performance in favour of Maxwell?

So this for example is a lie you think?
„Implements optimizations and bug fixes that provide increased performance for Kepler-based GPUs. “
-> http://us.download.nvidia.com/Windows/353.06/353.06-win8-win7-winvista-desktop-release-notes.pdf

I don't know, I was thinking I am gonna run a couple of test by myself using my 680

But how could you explain something like that otherwise?

http://i.imgur.com/w0PhEeQ.png

http://i.imgur.com/o4vCtEj.png

The delta is too large for being just some variance
 
Given that they both run the same frequencies, which vary greatly with notebook model and which I cannot tell from the screenshots, I'd say: Compare Maxwell and Kepler at the same ALU-counts, i.e. 980 Ti (2.816) vs. 780 Ti (2.880). In our performance-comparison (normalized and weighted average over 13 games in 3 resolutions), 980 Ti is about 40 percent faster than 780 Ti. Roughly normalized for clocks, there still would be 20 percent difference.

Further, there are architectural differences (shader export rate for example) that can amount to much more than that for example in pure fillrate tests.

I am not saying that Nvidia still is putting the same effort in optimizing for Kepler as for Maxwell (which I could not possibly assert anyway), but that I don't see very compelling evidence for your thread title's question.
 
If people report performance drops after driver "updates", then there is something. Those tables from the links are not compelling evidence? WOW!!!
 
Given that they both run the same frequencies, which vary greatly with notebook model and which I cannot tell from the screenshots, I'd say: Compare Maxwell and Kepler at the same ALU-counts, i.e. 980 Ti (2.816) vs. 780 Ti (2.880). In our performance-comparison (normalized and weighted average over 13 games in 3 resolutions), 980 Ti is about 40 percent faster than 780 Ti. Roughly normalized for clocks, there still would be 20 percent difference.

Factoring the peak compute performance, the difference is right there: the cut down GM200 is ~27% ahead of GK110 per clock cycle.
 
I've noticed a lot of chatter about this on several forums over the last few weeks. What's interesting is that there are a lot of people saying that they are going to boycott nVidia etc but what I haven't seen is proof that performance has degraded on Kepler cards.

I'm not into conspiracy theories but this has the feel of a widespread smear campaign. Hopefully a respected site picks up this story and does a comparison of the past few driver releases to confirm if/how Kepler performance has changed.
 
The cruel irony of this is that it makes Rebrandeon look like a strategy that pays dividends.

Would gamers really mind getting a rebranded chip if it means driver support will remain competent?
Now that's a new facet of the conversation that nVidia let slip.
 
I haven't seen is proof that performance has degraded on Kepler cards.

eM2Tzvn.jpg


cOn9tYq.png
 
Last edited by a moderator:

As with trinibwoy I don't like conspiracy theories but there's a hell of a lot of very compelling evidence there that's hard to ignore.

Without doubt, Kepler is performing worse relative to Maxwell in newer titles, that seems beyond debate. The real question is why? It seems the following suggestions are being made:

  1. NV is spending more time on Maxwell driver optimisations than Kepler driver optimisations. - This seems likely and is to be expected
  2. NV is deliberately hobbling Kepler performance through the driver to boost Maxwell sales. - This seems very unlikely as it harms their brand image overall. That said it's possible that where optimisations for Maxwell could harm Kepler performance then Maxwell is given presidence - which in a way amounts to the same thing and could explain the results in your second link
  3. Newer games are more geared towards GCN thanks to the console connection and thus perform generally better on AMD cards than Nvidia cards (Kepler or Maxwell), but in turn, Maxwell is more able to take advantage of the GCN optimisations than Kepler. - This is an interesting one and I guess could certainly be a real factor. Obviously AMD ARE performing better relative to NV in recent titles and the console connection seems to be an obvious reason for that. Given that Maxwell is more forward looking (like GCN) than Kepler, especially in how it handles async compute, then it's entirely possible those console optimisation are paying greater dividends on it vs Kepler
Personally I think there's an element of all 3 factors at play (to varying degrees depending on the game) with the exception of any deliberate work by NV to hobble Kepler performance (beyond a general apathy towards optimising it).
 
(beyond a general apathy towards optimising it)

This is an era where getting timely driver optimizations for the latest high-profile game is crucial for performance and stability, and nVidia is obviously aware of this.
The decision to deny the manpower for Kepler optimizations on the software level was calculated and made on purpose, no matter how you look at it.
 
Nothing out of ordinary. The GTX 960 tested here actually has ~25% more pixel fill-rate than the vanilla GTX 780 and looks like this particular game isn't ALU or TEX bound that much.
http://www.techspot.com/review/1006-the-witcher-3-benchmarks/page3.html

The 780 Ti is ~25% faster than the non-Ti. Their difference in fillrate is very small (same amount of ROPs, ~1,5% difference in clocks) whereas the difference in ALU and TMU count is no less than 25%.
So no, the fill-rate bottleneck excuse won't stick.




So all this fuss is over one game? Are people reporting the same issue on other games too?

Yes, it's been a rather consistent result among AAA game releases in 2015.

tVSRZ9A.png
FWAt2EY.png





Would also be interesting to see how Fermi does relative to Kepler in TW3.

This guy is claiming 30 FPS in 1200p using medium settings using a GTX 580, which is about the same as a Geforce 660 (GK106) according to Techreport.
It's certainly less shocking than seeing the 780 being beaten by / equalled to a chip with half the amount of transistors and using the same process node.
 
The 780 Ti is ~25% faster than the non-Ti. Their difference in fillrate is very small (same amount of ROPs, ~1,5% difference in clocks) whereas the difference in ALU and TMU count is no less than 25%.
So no, the fill-rate bottleneck excuse won't stick.
GTX 780 -- 32 fragments/clk
GTX 780 Ti -- 40 fragments/clk

Factor in the mild core clock difference (stock SKUs) and it scales just as expected.
 
If there is something to this, maybe it's temporary because the imminent release of Windows 10/DirectX12. Maybe they just don't have the resources to dedicate to both the new DX version and old hardware(on old DX version) at the same time. Hopefully.
 
GTX 780 -- 32 fragments/clk
GTX 780 Ti -- 40 fragments/clk

Source?
We have 12 SMX's enabled for 780 and 15 for the 780 Ti.
Both have all 48 ROPs and L2 cache enabled.
How do these "fragments/clk" (which I'm assuming it's the OpenGL equivalent to the old "pixel shaders" nomenclature) numbers associate with the number of active SMX and/or ROP units?
 
Those are synthetic benchmarks.
If your theory of Witcher 3 being solely bottlenecked by the fillrate presented in synthetic benchmarks was true, then the GTX 680/770 cards would be faster than the 780, which is not.

Besides, these aren't huge resolutions. This is 1080p, so fillrate requirements shouldn't be exactly stellar.
Plus, it doesn't explain the Kepler's performance deficit on Project Cars and Arkham Knight.
 
Back
Top