Is Nvidia deliberately downgrading Kepler performance in favour of Maxwell?

@elect it could simply be that the AMD hardware is more forward looking; I believe that has been noted elsewhere on this site. From personal experience my X1900XT really did well compared to it's competition in games released in the next few generations of hardware. Of couse one might attribute that to nVidia gimping their own hardware on purpose, but it seems unlikely.

Its easily possible to cause different kinds of bottlenecks on different games using the same engine. Add some more complex effects here, optimize something else there. The loads could be quite different.

Indeed, but Maxwell does seem to scale better in newer games than Kepler so if that's not a result of driver quality disparity between the 2 architectures then it begs the question - what you GCN and Maxwell have in common and that Kepler lacks that makes them better at running newer games. This is a really interesting question to answer IMO but would probably need dev input.
 
Indeed, but Maxwell does seem to scale better in newer games than Kepler so if that's not a result of driver quality disparity between the 2 architectures then it begs the question - what do GCN and Maxwell have in common and that Kepler lacks that makes them better at running newer games. This is a really interesting question to answer IMO but would probably need dev input.

It could just be that NVIDIA devotes more resources to driver optimization for Maxwell than Kepler, which would be understandable and quite different from sabotaging Kepler, as some people claim.
 
Indeed, but Maxwell does seem to scale better in newer games than Kepler so if that's not a result of driver quality disparity between the 2 architectures then it begs the question - what you GCN and Maxwell have in common and that Kepler lacks that makes them better at running newer games. This is a really interesting question to answer IMO but would probably need dev input.

Nothing common in particular between Maxwell and GCN. Maxwell's job was to redo the perf/watt gains that Kepler did to Fermi, without the need for a new process node. While Kepler was a new design paradigm of denser and simpler compute logic on a new process node at the same time, Maxwell took the knife and the ruler and trimmed all the fat overhead and re-shuffled the internal organization in more streamlined fashion -- do more with less on an already mature and well utilized process. GCN pretty much lived on incremental feature tweaks throughout the 28nm life span.

p.s.: AFAIK, Jensen mentioned last year that Volta would be based on Maxwell design. That would give some clue of how much importance NV places in this particular ISA now and for the future.
 
Last edited:
Nothing common in particular between Maxwell and GCN.

Maxwell and GCN have one thing in common. No dependence on ILP. In the absence of ILP Kepler's peak throughout drops by 33%.

All the evidence posted so far points to one simple conclusion - Maxwell is more efficient than Kepler in new games. This can be due to inherent architectural advantages or lack of targeted optimization for Kepler. Either way there isn't any evidence of some nefarious plot to sabotage Kepler based parts.
 
Maxwell and GCN have one thing in common. No dependence on ILP. In the absence of ILP Kepler's peak throughput drops by 33%.
Maxwell is still dispatching two instructions from a warp scheduler, just like Kepler and GF104/114. Kepler's low sustained throughput is due to insufficient register bandwidth, not the dual issue.
 
Nothing common in particular between Maxwell and GCN. Maxwell's job was to redo the perf/watt gains that Kepler did to Fermi, without the need for a new process node.

What about compute capability? I understand Maxwell is better at this than Kepler. Is it possible that the consoles are making greater use of compute which we are seeing both GCN and Maxwell taking advantage of while Kepler cannot?
 
What if they were controlling the tessellation factor in the driver for Maxwell cards, but not Kepler? Not only would this improve performance, but might also reduce power usage.
Not saying they are doing this, just a thought on what's possible.
 
It is like this - one generation ok, the next sucks, then again one generation ok.

Example - compute with Fermi OK, compute with Kepler NOT OK, compute with Maxwell - somewhat relatively good.

Please don't tell any of these to the 1,000s of machine learning applications developers . They might fall into despair knowing they are only using "NOT OK" or only "somewhat relatively good" hardware.
 
One thing I don't get is why newer games that don't have gameworks seem to still work well on kepler while gamework games work poorly on everything except maxwell.

Yes, as far as Gameworks titles are concerned, all of the high-profile games (except for Far Cry 4) have had some controversy related to performance and/or stability:
- Assassin's Creed Unity: royal clusterfuck
- The Crew: weird 60 FPS cap (was this ever removed?)
- Witcher 3: Sub-pixel polygons on hair/fur for tesselation sabotage more realistic simulation
-
Project Cars: Crappy performance on everything that isn't Maxwell
- Batman Arkham Knight: royal clusterfuck
- Borderlands Pre-Sequel: same engine as previous titles and like those, only PhysX is used. Also meh game.
- Watchdogs: royal clusterfuck

That's very telling.
 
Last edited by a moderator:
One thing I don't get is why newer games that don't have gameworks seem to still work well on kepler while gamework games work poorly on everything except maxwell.

Thats an interessant question, but at the same time, the non gamework titles are offtly games who are not under the Nvidia program ( or maybe just as standard logo appearance ), and they dont have the possibility to developp the driver or collaborate to developp the games as much they can, or want like with Gameworks title.

The question is maybe more for gamework titles, are they only optimizing the driver extremely well for maxwell, forgetting older gpu's, or is the code is specially favorizing maxwell over kepler ?...

But for what i have see, gameworks features seems not to be directly the cause of the low performance of old gpu's ( not like with AMD GPU's in general, who see their performance tanks at the moment you enable gameworks features ).

As for the dramatic states of gameworks titles when launched ( and the list start to be long and all in follow ) , i think the problem is not gameworks features themselves, but something seems complicate the developpement of the games.. maybe its not a so good thing to get some part of engine who are not fully in the hand of the developpers ? or is it a question of added work ? But this is not the right thread for discuss it..
 
Last edited:
Yes, as far as Gameworks titles are concerned, all of the high-profile games (except for Far Cry 4) have had some controversy related to performance and/or stability:
- Assassin's Creed Unity: royal clusterfuck
- The Crew: weird 60 FPS cap (was this ever removed?)
- Witcher 3: Sub-pixel polygons on hair/fur for tesselation sabotage more realistic simulation
-
Project Cars: Crappy performance on everything that isn't Maxwell
- Batman Arkham Knight: royal clusterfuck
- Borderlands Pre-Sequel: same engine as previous titles and like those, only PhysX is used. Also meh game.


That's very telling.

Unity and Batman are the only titles with actual issues and neither have been found to be gameworks specific.

Project cars runs just fine on other cards and the other issues you raised wouldn't even qualify as petty.
 
Maxwell can't dual-issue ALU ops.
I stand corrected:
"The power-of-two number of CUDA Cores per partition simplifies scheduling, as each of SMM's warp schedulers issue to a dedicated set of CUDA Cores equal to the warp width. Each warp scheduler still has the flexibility to dual-issue (such as issuing a math operation to a CUDA Core in the same cycle as a memory operation to a load/store unit), but single-issue is now sufficient to fully utilize all CUDA Cores."
 
Yes, as far as Gameworks titles are concerned, all of the high-profile games (except for Far Cry 4) have had some controversy related to performance and/or stability:
- Assassin's Creed Unity: royal clusterfuck
- The Crew: weird 60 FPS cap (was this ever removed?)
- Witcher 3: Sub-pixel polygons on hair/fur for tesselation sabotage more realistic simulation
-
Project Cars: Crappy performance on everything that isn't Maxwell
- Batman Arkham Knight: royal clusterfuck
- Borderlands Pre-Sequel: same engine as previous titles and like those, only PhysX is used. Also meh game.


That's very telling.

Don't forget to add Watchdogs to that list. Even FarCry 3 had some issues on some Nvidia cards (Total Biscuit on Youtube had frame pacing issues and noticeable hitching issues on his Nvidia cards).

Regards,
SB
 
Don't forget to add Watchdogs to that list. Even FarCry 3 had some issues on some Nvidia cards (Total Biscuit on Youtube had frame pacing issues and noticeable hitching issues on his Nvidia cards).

Regards,
SB


Huum, Farcry3 was not an Nvidia games if i remember well, was under the Gamevolving program (sometthing who have change for FC4 ), this said, at launch the stutter issue was nearly for everyone ..
 
Back
Top