Now don't freak out. I'm going to give all of the results to you at once.
No, they're not simple bar or line graphs that a lab monkey could read, but I'm confident
that you can beat a monkey handle it.
There has been some noise lately about whether SLI is living up to the
marketing claims of "...an SLI configuration will come close to doubling your 3D
acceleration power". Technically, NVIDIA's claims were "up to 1.83 times" or
83%. Anyone that's been around computers (and video cards specifically) though,
knows that marketing claims are generalizations. There are an infinite amount
variables at play in the real world: CPU, motherboard, memory, BIOS settings,
drivers, games, game settings, etc. So, can one realistically expect an 83%
performance gain all the time? Of course not.
NOTE: The increases or decreases are relative depending on the directionality
you apply to the differences. For example, the CodeCreature
results for a single card show a 49.9% decrease from 0xAA/0xAF (4514) to
4xAA/8xAF (2262). If you start with the 2262 score though there is a 99.6%
increase to 4514. If that went over your head then you can just ignore the
change percentages or you can play around with the provided calculator below.
The default Forceware SLI profiles were used in the benchmarks below, no
forcing of AFR or SFR was performed. The game benchmarks were all run at
1600x1200 resolution. Anything less than 1280x1024 is pretty pointless when comparing SLI benefits
from both a technical and (economically) practical perspective. It's been stated
many times but is worth repeating. If you only game at 1024x768 resolutions
then going to SLI makes no sense. Only hi-res gamers and high AA/AF users need
UPDATE: I should mention that I ran each timedemo four times and took
the average of the last three to come up with the results below for both
Half-Life 2 and Doom 3. Also, AA/AF was set to "Application Controlled" in the
Forceware Driver Control Panel for both games.
SLI Comparison: The default benchmark was run and you can see
a respectable 73.6% increase when SLI is enabled.
SLI Comparison: The gains here with SLI are
decent and quite a reduction shows up when AA/AF are enabled.
AA/AF Comparison: There's naturally a performance hit when
moving from 0xAA/0xAF to 4xAA/8xAF and the hit is substantial here
even with SLI enabled.
UPDATE: The demo1 results above are
invalid. More info here.
It has come to my
attention that the Doom 3 results are
This article helps to shed some light on
this issue. This was an honest mistake on my part and I sincerely
apologize for any confusion.
UPDATE:One of my fellow staff members, Brian Gray (aka
BrianG, aka HalcYoN), thought it would be interesting to reveal the
point at which the memory bandwidth restriction of a single 6600GT
card comes into play compared to dual 6600GTs in SLI mode. The Doom
3 trdelta1 timedemo proved to be a good case study for this.
In order to illustrate the memory bandwidth restriction points it is
necessary to include 1024x768 and 1280x1024 results alongside the
1600x1200 results. NOTE: The 1600x1200 results below are
identical to what is found above and simply copied here to keep this
particular comparison together.
I'll let Brian elaborate on these results in more detail as he may
be putting together an op-ed on how some folks are losing sight of
the real advantage of SLI.
SLI Comparison: I used three of the five
timedemos that are included in the
HardwareOC Half-Life 2 Benchmark v1.3. I did not use the
benchmark program though. I ran the timedemos myself from the HL2
console. Also, I ran benchmarked with Steam in offline mode and with
my network connection disconnected.
Half-Life 2 is a classic
example of a game that kind of shrugs at you when you throw an SLI
setup at it. With hoc_coast1 you can see a case where the CPU
overhead involved with SLI shows its teeth a bit by actually
dropping performance a bit. The rest of the 0xAA/0xAF results show
some modest gains for the other timedemos but the the 4xAA/8xAF are
AA/AF Comparison: The changes
with SLI disabled and enabled are about what one would expect.