7950 or 670?

[lopri]

I've found an unexpected workaround on 2D-3D frequency issue on HD 7950. At least for Flash.

My browsers of choice are Chrome and Opera. Those who've used Opera on a regular basis know it offers so much tweakability. It currently supports most GPU-intensive web page rendering among the browsers. (possible exception is Safari, and I haven't used Safari for so long so I do not know how it handles web pages)

Opera now can render web pages almost completely using GPUs, via WebGL. (Tutorial -> http://my.opera.com/desktopteam/blog/2012/04/20/update-on-hardware-acceleration-in-opera-12) What it does is predictable: Frees CPU cycles, provides better graphical performance as well as text anti-aliasing. (but it eats up video memory like there is no tomorrow)

When I enabled this feature in Opera, the symptom of defaulting to 2D frequency under any situation is gone. I have just found it out and tested only for Flash, and I tentatively confirm. "Tentatively" because I only tested 4 video clips so far. The card clocks to whichever task that are occurring on the screen at that moment and that is more demanding.

This still doesn't fix the off-line HD playback on most media players I have tested. (WMP, MPC-HC, Quicktime, Adobe Media Player, etc.) For those, it still defaults to its performance 2D frequency. (500/1250)

It's still an interesting found, I thought. Those who use Opera might give it a try.

 

[lopri]

Note: This feature of Opera doesn't get along with unstable overclock. Like NOT AT ALL. (Both for GTX 670 and HD 7950)

 

[RussianSensation]

Why not disable hardware acceleration for GPU in flash and let the CPU handle it? CPU usage is very small and the flash clock issue is fixed. Modern CPUs can handle 1080P video without a problem.

 

[lopri]

I agree that's a better way to go as far as Flash is concerned. I just thought it was interesting because it wasn't expected.

 

[lopri]

It is quite strange to see Tahiti and GK114 have equal theoretical performance per clock frequency, I have to say. This was something I haven't seen for a long time. Since G80, GeForce had a clock generator for independent shader clocks (usually x2 of core) so that was obviously one part of the (un)equation.

I think this is really the bottom line of similar 3D performances between these two.

 

[Silverforce11]

Yeah its quite amazing that two very different engineering team and companies working on different design goals arrive with two completely different architectures, which somehow manage to perform so close to each other.

 

[RussianSensation]

I think this is really the bottom line of similar 3D performances between these two.

Often times, you can't really use theoretical pixel/texture fill-rates in GPU-Z to conclude that it's the main reason the cards perform similarly. For example comparing GTX580 vs. 6970 (what's more important a lot more pixel fill-rate or texture fill-rate? Depends on the game). That's because the theoretical # often has little to do with actual throughput the card can achieve in the real world.

For example, even with the same number of ROPs and a similar theoretical fill-rate performance limit (29.6 vs 28.16), 7970 is pushing 51% more pixels than 6970 is. Without checking actual real world pixel/texture fill-rate performance of cards, looking at the GPU-Z # to arrive at any conclusions can be misleading:

The fact that it worked out this time could be a mere coincidence (i.e., both GK104 and Tahiti have similar efficiency in the real world for pixel/texture fill-rate). Even then you have to be careful since there are different texture depth and performance can vary based on that (INT8 vs. FP16 texture filtering):

Also, theoretically speaking, GTX670/680 should easily beat Tahiti in games that use a lot of GPU cloth physics or particle effects:

While AMD would tend to do better with a lot of parallax occlusion mapping (Crysis 3):

Since games often use a combination of these features, each card runs into its own bottleneck somewhere, and that's not even touching tessellation/geometry performance and memory bandwidth efficiency which is critical for higher resolutions and higher AA modes. So I think it's too simplistic to just compare texture and pixel fill-rate performance between the 2 brands since not all TMUs and ROPs are created equal (Cayman vs. Tahiti ROP comparison explains this). You can't really compare 32 ROPs of one GPU and 32 ROPs of another GPU without doing further testing even if their theoretical #s match on paper.