The G80 delivers stellar increases in IQ/perf on DX9 games over G7x, but many people will reason that the true power won't be unlocked until DX10 games arrive, which are along way off, and thus, in the meantime, the card will only be a decent upgrade to topend G7x/R5xx.
However, some thinking on the crapper has convinced me that the G80 will kick start the physics market long before even DX10 games arrive, in a way that Ageia can only dream of. We've seen hints of the G80 physics power in the "smoke box" demo which runs real time Navier-Stokes computation fluid dynamics calculations on a very fine grid concurrently with a raytracing/raymarching rendering app, showing off single-GPU usability of physics.
Whereas Ageia was coupled to Novodex on their HW (which IMHO was a Physics *decelerator*), NVidia is pushing Havok, which has the majority of the middleware physics marketshare. This brings up the possibility of a recompile/relink/patch of existing games being able to take advantage of GPU physics, perhaps with increased "Effects Physics" loads in the beginning.
NVidia's "Quantum Physics Engine" seems to be a low-level port of Havok to NVidia's CUDA architecture, which seems to repurpose L1/L2/GPR/onchip storage into scratch-pad RAM ala CELL, albeit with 8 times the scalar MADD power of a single CELL, as well as a much faster memory bus and better TLP design/branching performance. If the QPE can be dropped into existing Havok apps with minimal upgrade work, it could show off immediate benefits/killer demos in a way that Ageia never could, in that one doesn't need to buy a "PPU only" card.
Besides the fact that for the first time, a GPU card is on the market which may be able to accelerate physics reasonably well (and indeed, accelerate more than effects physics with no collision detection),
SLI becomes even more worthwhile, even for low-end G8x cards as a way to add-on PPU power later. It's unlikely that people with SLI 8800GTXes will encounter resolution limits in the near term on DX9 games, and higher SLI AA levels have diminishing returns (unless we're talking about getting into temporal supersampling, etc) So there is less pressure to own an SLI 8800GTX given its performance on rendering existing games. However, QPE performance would be a major selling point, if you could cut down on context-switch time and memory/cache thrashing (between GPU rendering and CUDA mode) by adding a second card, as well as maximizing physics performance.
Leaving aside Folding@Home stuff, I also think we will see this go beyond Newtonian dynamics and branch out into audio processing (resurrect Aureal Wavetracing please), as well as the mundane, such as boosting multimedia apps (Adobe) and even server apps (think SSL acceleration)
But my main point is, Ageia was too overpriced for a single-purpose solution, as well as IMHO, being unable to compete with NVidia/ATI from an engineering perspective (process, memory bus, design, et al). Physics on previous-gen non-DX10 chips seems to be limited due to architecture limitations. Stream out helps, but is less useful than having the additional ability (in addition to streamout) for shared general purpose random access memory + synchronization primitives (G80). Now granted, the total amount of memory is small, which means it won't be a simple cakewalk to get it working, but the potential is alot higher.
Thus, it seems to mean that 2007 could be the year of GPU physics acceleration. True ground-up DX10 games (as opposed to quickies which use some DX10 features) might not arrive until 2008. However, middleware upgrades to use GPUs may bridge the gap in which the extra (and in some cases hidden/unexposed-by-DX10) GPGPU power of these cards may be used to upgrade existing engines with nice add-on features (remember how various games got EAX/Aureal/MMX/SSE "patches"?)
Now, it is true that 2007 will also be the year of quad-core CPUs and ubiquitous dual-core chips, and that means games with increased physics workloads will also run on these systems too, but NVidia is claiming 10x physics performance vs a 2.66 Ghz Core Duo already, so whatever additional CPU power these machines have will only help make the QPE setup quicker, load balance with the QPE (CPU + GPU physics acceleration) or reduce CPU limits elsewhere.
A counter argument would be that G80 ain't gonna run Havok much faster than a Core 2 Duo, that perhaps Havok collision detection won't run on it, or will run, but won't have a stellar speedup. Or, that a 10x speedup given n^2 or n^3 scaling won't lead to that much of a visual difference. I think those are legitimate points to consider.
Finally, there's the aspect of a "physics only G80". Why? When I just explained how consumers will be offput by single-use solutions. I believe that this may come to fruition, and will be a way for "non-SLI" mainboard owners to boost performance. This would be a cutdown G80 core with smaller memory bus, single-slot card installable in a secondary PCI slot. During the NVidia G80 launch, they held up an enthusiast 680i mainboard with *FOUR* Nvidia cards in it. Two were 8800GTXs, dual slot consuming cards. The other two were thin single-slot cards presumably sharing a third space where you can plug in either another dual-slot 8800GTX, or 1 or 2 single-slot GPUs. There are two potential markets for this depending on price: 1) if they can get reasonable performance on a cut-down G8x for under $100, it might sit well with mainstream users. 2) the SLIzone overclocking "I spent $2000 on my gfx system alone" crowd.
Ok, enough bathroom crapper theories.
However, some thinking on the crapper has convinced me that the G80 will kick start the physics market long before even DX10 games arrive, in a way that Ageia can only dream of. We've seen hints of the G80 physics power in the "smoke box" demo which runs real time Navier-Stokes computation fluid dynamics calculations on a very fine grid concurrently with a raytracing/raymarching rendering app, showing off single-GPU usability of physics.
Whereas Ageia was coupled to Novodex on their HW (which IMHO was a Physics *decelerator*), NVidia is pushing Havok, which has the majority of the middleware physics marketshare. This brings up the possibility of a recompile/relink/patch of existing games being able to take advantage of GPU physics, perhaps with increased "Effects Physics" loads in the beginning.
NVidia's "Quantum Physics Engine" seems to be a low-level port of Havok to NVidia's CUDA architecture, which seems to repurpose L1/L2/GPR/onchip storage into scratch-pad RAM ala CELL, albeit with 8 times the scalar MADD power of a single CELL, as well as a much faster memory bus and better TLP design/branching performance. If the QPE can be dropped into existing Havok apps with minimal upgrade work, it could show off immediate benefits/killer demos in a way that Ageia never could, in that one doesn't need to buy a "PPU only" card.
Besides the fact that for the first time, a GPU card is on the market which may be able to accelerate physics reasonably well (and indeed, accelerate more than effects physics with no collision detection),
SLI becomes even more worthwhile, even for low-end G8x cards as a way to add-on PPU power later. It's unlikely that people with SLI 8800GTXes will encounter resolution limits in the near term on DX9 games, and higher SLI AA levels have diminishing returns (unless we're talking about getting into temporal supersampling, etc) So there is less pressure to own an SLI 8800GTX given its performance on rendering existing games. However, QPE performance would be a major selling point, if you could cut down on context-switch time and memory/cache thrashing (between GPU rendering and CUDA mode) by adding a second card, as well as maximizing physics performance.
Leaving aside Folding@Home stuff, I also think we will see this go beyond Newtonian dynamics and branch out into audio processing (resurrect Aureal Wavetracing please), as well as the mundane, such as boosting multimedia apps (Adobe) and even server apps (think SSL acceleration)
But my main point is, Ageia was too overpriced for a single-purpose solution, as well as IMHO, being unable to compete with NVidia/ATI from an engineering perspective (process, memory bus, design, et al). Physics on previous-gen non-DX10 chips seems to be limited due to architecture limitations. Stream out helps, but is less useful than having the additional ability (in addition to streamout) for shared general purpose random access memory + synchronization primitives (G80). Now granted, the total amount of memory is small, which means it won't be a simple cakewalk to get it working, but the potential is alot higher.
Thus, it seems to mean that 2007 could be the year of GPU physics acceleration. True ground-up DX10 games (as opposed to quickies which use some DX10 features) might not arrive until 2008. However, middleware upgrades to use GPUs may bridge the gap in which the extra (and in some cases hidden/unexposed-by-DX10) GPGPU power of these cards may be used to upgrade existing engines with nice add-on features (remember how various games got EAX/Aureal/MMX/SSE "patches"?)
Now, it is true that 2007 will also be the year of quad-core CPUs and ubiquitous dual-core chips, and that means games with increased physics workloads will also run on these systems too, but NVidia is claiming 10x physics performance vs a 2.66 Ghz Core Duo already, so whatever additional CPU power these machines have will only help make the QPE setup quicker, load balance with the QPE (CPU + GPU physics acceleration) or reduce CPU limits elsewhere.
A counter argument would be that G80 ain't gonna run Havok much faster than a Core 2 Duo, that perhaps Havok collision detection won't run on it, or will run, but won't have a stellar speedup. Or, that a 10x speedup given n^2 or n^3 scaling won't lead to that much of a visual difference. I think those are legitimate points to consider.
Finally, there's the aspect of a "physics only G80". Why? When I just explained how consumers will be offput by single-use solutions. I believe that this may come to fruition, and will be a way for "non-SLI" mainboard owners to boost performance. This would be a cutdown G80 core with smaller memory bus, single-slot card installable in a secondary PCI slot. During the NVidia G80 launch, they held up an enthusiast 680i mainboard with *FOUR* Nvidia cards in it. Two were 8800GTXs, dual slot consuming cards. The other two were thin single-slot cards presumably sharing a third space where you can plug in either another dual-slot 8800GTX, or 1 or 2 single-slot GPUs. There are two potential markets for this depending on price: 1) if they can get reasonable performance on a cut-down G8x for under $100, it might sit well with mainstream users. 2) the SLIzone overclocking "I spent $2000 on my gfx system alone" crowd.
Ok, enough bathroom crapper theories.
