HAVE YOU EVER WONDERED how much real performance gain one would get from having two or three graphics cards in parallel? After all, Nvidia SLI and ATI Crossfire have been around for years now, and the software titles must have taken at least some advantage of the opportunity. However, unlike the multi-core parallel CPU situation where there are always more tasks to throw at the extra cores in modern OS environments, the benefits of multi-GPU setups are limited to heavy graphics workloads. And recently, of course, some computational workloads as well.
Don't expect anywhere near twice the performance here for your typical game or even engineering 3D simulation. However, the benefits of parallelism are there, and now it seems to scale acceptably well on three or even four GPUs.
With the new Catalyst 10.3 driver, AMD has now separated the Crossfire application optimisation updates into files that can be updated more often than the regular monthly driver cycle, if required. How far does it go right now?
Here we have the ultimate system in terms of CPU, memory and I/O, the EVGA W555 SuperRecord 2 mainboard pre-release version with two six-core Intel X5680 Xeon processors running at 4GHz sharp, with six channels of Kingston HyperX DDR3-2000 memory running at DDR3-1600 CL 6-6-6-16.
The board has of course four full x16 PCIe v2 slots through the two Nvidia Nforce 200 bridges. Yes, those come out of only dual x16 PCIe v2 lanes from the Tylersburg chipset, but the internal broadcasting capability of the Nforce chipset may help performance in some cases, too.
Then, the Intel X25-M 160GB SSD feeds the system with data, while the Corsair HX1000 PSU provides the power. So, there are no speed bumps here driving any GPU combination, I guess.
Now, the GPU side of the equation was taken care of by trusty Asus. On one side, its HD5870 cards have the unique Voltage Tweak feature for easier overclocking, and up to 930MHz GPU speed can be achieved. Asus' custom design cards, like the HD5830 OC edition, can go even higher - look for our coverage on that soon.
Here I put one, two and three of these HD5870's in a parallel Crossfire combination. In addition, a pair of the HD5970 dual GPU cards are also on hand for up to quad GPUs testing, which you'll see in Part 2 of these benchmarks soon.
However, I didn't do any GPU overclocking in this case, as it is usually more complex and less rewarding in terms of actual results than a reasonable CPU performance adjustment. Therefore, it's default GPU speeds here all the way.
So, how was the 3Dmark Vantage scaling? Here we have the HD5870 Performance results in one, two and three GPU setups.
Then, the HD5870 Extreme results in the same configurations.
Nice numbers and pretty decent scaling there. Remember, here we have newer numbers than we've seen before, as we used the newer Catalyst driver.
There's another benchmark that has attracted gamers' attention recently, Heaven 2.0 from Unigine. A superb DX11-enabled test with tesselation and a variety of options plus a free walkthrough that's missing in 3Dmark tests, brings the most out of graphics cards. And, it has a Linux version supporting OpenGL 3.0 natively. With a vastly different engine and dataset, how does it scale in Crossfire?
We went ahead with two settings. The first is the default high-end DX11 test with tessellation and shaders on high.
The second is a DX10 test with medium shaders and no anisotropy.
Finally, here are the results across all combinations.
Not much scaling here. Maybe the Catalyst 10.4 driver or an intermediate Crossfire update profile would solve this, but generally such a test should scale as well as the GPU tests from 3Dmark Vantage.
In particular note that the DX10 mode tests with lesser settings scale much better than the DX11 tests, where the dual card setup is in fact slower than the single card configuration, and the triple card setup couldn't even complete the DX11 run no matter how many times I tried.
In summary, it really depends on what you run. Just like 3Dmark Vantage, there are quite a few games and 3D graphics applications that do make use of Crossfire and SLI alike. Some others, like Heaven 2, will have to wait a little longer, but hopefully not until we're all in, er, heaven.
In the meantime, single and double HD5970 results will be upcoming soon with the same tests. µ
Uses 20 percent less power than traditional systems
It's becoming more prevalent in car research and development
Sign up for INQbot – a weekly roundup of the best from the INQ