Update! I've uploaded stock-speed i5 760 results with 7970 and 7970/CF to my site (except for 3DMark,
doing that shortly, ie. 3DMark06, Vantage, 3DMark11, Firestrike, etc.) I don't yet have pages for most of
the 3DMark data I've accumulated, but I'll include the links here when they're ready; atm my site only has
a page for the '06 results). See:
http://www.sgidepot.co.uk/sgi.html#PC
Thus, check the data for Stalker COP, Unigine, X3TC, Call of Juarez, Far Cry 2 and PT Boats.
NB: remember the 7970 GHz Edition is basically the same as a 280X. I have two GHz Edition
7970s, one runs at 1050MHz, the other normally at 1150 but I force it to 1050 in CF so it
matches the other card.
Results are much as I expected with a 7970. At low res/detail, CPU speed does make a difference,
but the fps are so high it doesn't matter anyway. At high res/detail, the gap narrows or vanishes
completely (sans dodgy Unigine/Stalker results because of the DX9/CF issue), yet performance is
still good overall.
DX10/DX11 tests show the expected scaling, with Stalker @ 2560x1440 high-detail 7970/CF very
close to the speed of the same 7970/CF with a 5GHz 2700K. This is what I meant about the way
many games become less & less bottlenecked by CPU power as the resolution/detail/etc. increases.
The gap is a little wider for Ungine Valley @ 1920x1080 DX11, so I expect an oc'd i5 to help there
(notice how the gap shrinks when moving to 2560x1440). Some newer games can behave a bit
differently, but that's less common.
FC2 at 1920x1200 does show a gain from two 7970s even with the i5 @ stock speed, but the gap
between it and the 2700K suggests this a typical example where having the i5 oc'd would help, though
note (as above) the performance is so high one wouldn't really notice. At 2560x1440 the CF scaling gap
increases, which is good, and the divide between the i5 and 2700K sihrinks somewhat, but it's still
significant. Oc'ing the i5 would help, but again the fps is so high one would not notice in practice, except
of course for the larger degree of stutteriing which can occur with multiple GPUs for some tests (this
issue is very noticeable for the high-res Unigine Heaven/Valley tests).
Call of Juarez at low res shows a gap dependent on CPU power, but as above the absolute speed is
so high it doesn't matter. Move up to 1920x1080 though and everything changes. Now the CF scaling
is working very well; the i5 isn't far behind the 2700K, and more interestingly their results for just one
7970 are almost identical. COJ is the classic case where increasing res & detail very quickly moves the
bottleneck onto the GPU. Alas I don't know how to run this test beyond 1920x1200 (doesn't give me
the option of testing any higher), so I don't yet know how performance would scale up to 2560x1440.
PT Boats is an odd one, it seems to be CPU-limited at just about any resolution, perhaps because it
employs lots of physics calculations which normally would be accelerated via PhysX on an NVIDIA card,
but of course with a 7970 it has to be done by the main CPU.
Tropics seems to have CPU-related issues at low res or with DX9, but these factors begin to fade
at HD res and higher. Indeed, at 2560x1440 with DX10, the i5 760 beats the 2700K!
(TBH it's
basically exactly the same, ie. well within error margins, but the result is nonetheless amusing,
showing how the GPU-heavy nature of Unigine sometimes only reveals itself when under heavy load).
I think the Heaven results show some potential for an oc'd CPU helping out even at medium detail,
it will be interesting to see what happens with the oc'd setup.
One thing is clear though (and the 3DMark data will likely back this up when I have the numbers),
a P55 system is perfectly capable of exploiting powerful cards quite well, much more so than
most people would believe, and remember most P55 boards are only running with x8/x8 (similar
limited PCIe issue to modern Z97, though at least the newer chipsets are V3, so x8/x8 on a Z97
is akin to x16/x16 on an older board). I do have a P55 board which supports x16/x16 (and also
x8/x8/x8/x8; it's an ASUS P7P55 WS Supercomputer, employs PCIe switches in the usual manner),
but I've not done much with it yet, and I won't use it for most results because such data wouldn't
be representative of most P55 boards.
Anyway, I'll sort out the oc later this week and update the tables with oc'd numbers. Can't do it
sooner, got to do some work stuff first.
Hope this helps!
Ian.