Technical speculation- why has Kepler "aged badly"?

[ThatBuzzkiller]

Console optimizations will do that ...

AAA games will start using more compute shaders since that is the optimal path for consoles, local memory atomics are very slow on Kepler microarchitecture and although they vastly improved it with the Maxwell microachitecture it's still slower than the GCN microachitecture found in consoles in that regard and it's no secret that devs do a lot to minimize the impact of geometry heavy passes in their games due pre-Polaris microachitectures low geometry throughput so the one definitive advantage that Nvidia has isn't used much ...

The shaders get optimized for consoles and Nvidia keeps replacing their driver stack too so driver optimizations don't hold between each GPU microachitecture from Nvidia ...

It's a combination of all these things that will comparatively weaken Kepler's performance in newer games ...

 

[NTMBK]

Console optimizations will do that ...

AAA games will start using more compute shaders since that is the optimal path for consoles, local memory atomics are very slow on Kepler microarchitecture and although they vastly improved it with the Maxwell microachitecture it's still slower than the GCN microachitecture found in consoles in that regard and it's no secret that devs do a lot to minimize the impact of geometry heavy passes in their games due pre-Polaris microachitectures low geometry throughput so the one definitive advantage that Nvidia has isn't used much ...

The shaders get optimized for consoles and Nvidia keeps replacing their driver stack too so driver optimizations don't hold between each GPU microachitecture from Nvidia ...

It's a combination of all these things that will comparatively weaken Kepler's performance in newer games ...

I hadn't realised that GCN had such superior local memory atomics. Explains why Nvidia prioritised it in Maxwell!

 

[Innokentij]

Have a read : https://www.computerbase.de/2017-01/geforce-treiber-test/2/

 

[wilds]

I would also like to add that the initial release of Kepler had too little VRAM; shortening it's life expectancy.

GTX 670, 680, 770 all had 2 GB variants that should not have existed at that price range. The 780 and 780 ti 3 GB versions seemed adequate at the time, but it wasn't enough for a long life that 8 GB cards can ensure.

Fortunately Kepler is still very good at DX9 and older OpenGL titles. My tiny 650M still performs admirably with 1 GB of VRAM holding it back in newer titles.

I am curious to how many GTX 780 6 GB, OG TITAN and TITAN Black owners there are out there that are happy with their cards.