I've heard AMD owners talk about this compute stuff in discussions relating Nvidia Kepler VS the latest AMD cards.. As I've understood it, these functions are not needed for gaming and just adds extra power and heat to a gaming card.
Double precision is worthless for games but you can't clump all compute features as worthless. Compute isn't some magical term that keeps getting thrown around on our forum, but what it means is a GPU can perform a wide variety of computational work that normally it couldn't do efficiently otherwise. For that you need a complex dynamic scheduler, compute-focused GPU architecture, an entirely new way of designing the GPU around compute. In essence, the GPU becomes a general purpose processor, that happens to play games well. What we currently have are GPUs that can play games well and very little of anything else. The evolution of GPUs means it has to become more compute focused, whether NV likes it or not or they'll be left behind. NV knows this and its future architectures will be even more compute focused with shared address space, memory pre-emption, etc. Just because Kelper gaming card is neutured in compute doesn't mean that NV doesn't want to make their GPUs more general purpose, which is the whole point behind GPGPU initiatives NV started with G80.
This is why Tahiti XT is a much more general purpose GPU than Kepler GK104 is. For the first time since G80, NV was able to decouple its architectures into a pure gaming chip and a compute GPU for Tesla. They didn't do this with Fermi. The 2048 SPs are just what gamers look at but it's actually
32 Compute units, each having 64 SPs. Those compute units are fed by the
ACE command engines in Tahiti. Why is this important? Tahiti XT isn't just a chip with double precision fat but
the entire GPU was re-designed around compute first, graphics second. It shows a major change in thinking about what a GPU should be. So unless NV thinks compute is not the future of GPUs becoming more general purpose, they will need to follow at some point. My guess if Maxwell.
What do you think allows Tahiti XT to outperform GTX680 in Dirt Showdown, Sleeping Dogs and Sniper Elite V2? There is no magic sauce but AMD uses more efficient CUs to perform lighting passes and calculate HDAO/Anti-aliasing passes without the much heavier penalty a traditional GPU architecture would incur. When you refer to "compute stuff" let's separate double precision which you don't need for games from "compute" functionality that GPU can have to perform graphical effects. To perform non-traditional compute work on the GPU, you still need an architecture that can perform compute work effectively and that's why you have dynamic scheduler and additional units around stream processors telling them how to allocate the work. All this takes up extra transistor space. It's not just the double precision aspects.
NV got off very easy this round since almost no games used compute extensively. No one really knows how many games in 2013 (if at all) will use Compute to accelerate certain graphical aspects but if they do, NV can't just forget about "compute stuff" and close a blind eye to it or they'll be left behind using slower methods to say perform a lighting calculation and then get creamed just like in Dirt Showdown.
Dirt Showdown is a perfect example of what would happen if games started using
open-standard DirectCompute for more advanced lighting path. It's a fully open standard and NV cards tank at it because Kepler GK104 can't perform "compute stuff" well.
Not only that, but DirectCompute can be used to give us back the MSAA we loved. It is exactly DirectCompute path that allowed AMD to used Deferred Shading in the Leo Demo and which showed efficiencies of that technique without the limitations of Deferred Rendering Path + MSAA we now experience.
Can NV not care about "compute stuff" for its GPU architecture for another generation? Probably because it's not like compute in games is taking off and it's not like NV can redesign Kepler in 12 months - which is basically just Fermi enhanced and not much more. I just want to make a very clear point that "Compute" for GPUs (heterogeneous GPU computing) and "Double Precision"
do not mean the same thing. You don't need DP for games but you can use DirectCompute to accelerate graphical effects. If future games actually use DirectCompute, then compute performance of a GPU will become a huge factor in performance. It's a chicken and an egg scenario. Without next generation GPUs that can perform DirectCompute well, no one is going to develop games that use compute extensively.
BTW: Even a "crippled GK110" is faster than AMD's S9000...
Ya, that's because in the professional industry, software is probably 90% of the results. Until AMD gets developers to make extensions/alter code to take advantage of its architecture, it won't be faster using traditional code/programs.
Look at consumer applications like distributed computing or programs that use OpenCL compute that GCN architecture leverages. It absolutely creams Kepler in those apps, because those programs have been updated to support the latest OpenCL/double precision support (MilkyWay @ Home). Professional apps need a lot more validation and take time for these major changes to trickle down. You are not going to fire Photoshop CS6/
WinZip 16.5 and have GCN lay waste to a Core i7 3960 unless the developer adds extensions/programming language that can use the GCN's capabilities. That's why older versions of WinZip and Photoshop don't run any faster on GCN.
Without modern software, GCN has little to offer over Kepler in professional space. Develop an application that uses OpenCL or double precision compute and optimize it for GCN and it would be a lot more competitive. Since NV has what like 95% of the professional graphics segment, it also has a lot of experience of working on extensions for its GPUs and working closely with developers on its CUDA optimizations. I am not saying GCN will ever beat current Quadro or Tesla cards for professional apps but AMD has to start somewhere. GCN's more advanced architecture over Kepler for compute might not even matter since AMD has no $ to invest into professional GPU software support right now given the financial state of the company. NV will launch Maxwell in 2014 and by then it'll be entirely different ballgame, meaning NV could develop a superior compute architecture than GCN is.
In actual gameplay you can't see any difference. Maybe in strategy games you would.
