CPUs

While GPUs continue to become more sophisticated with each generation, and begin to look suspiciously more like massively parallel general-purpose processors than highly specialised shader processors, the humble CPU still has a massive effect on your PC's performance.

The latest processors have as many as 820 million transistors, a good 120 million more than the most complex GPU, the Radeon HD 2900 XT. CPUs also lead the charge when it comes to manufacturing process technology; Intel offers a full range of 45nm CPUs, while Nvidia and ATI still use 55nm and 65nm processes. Because of this, manufacturers have been able to dramatically increase the performance of their CPUs, adding more cores, more cache and new architectural features, while lowering power consumption and price. This time last year a quad-core CPU would set you back £400, but the same CPU can now be had for as little as £137.

These developments have made it hard to choose which CPU to upgrade to next, as so many models have flooded the market. Last year's megatest included 54 models, but this year, we had to test 70 models to ensure that we could review every CPU on the market. With so many CPUs available, there are duff models to be avoided, and others that are so good you might be tempted to buy a few to build a folding farm. Let Custom PC guide you through the CPU minefield to find the perfect processor.

In the same way that our method of testing PCs has evolved over the years, so too has the way in which we measure performance. For this year's CPU megatest, we devised a suite of nine tests ranging from editing photos and encoding videos to a massive nuclear war in Supreme Commander.

First off, we ran our Media Benchmarks 2007, which you can download from www.custompc.co.uk and run on your own PC. The Media Benchmarks comprise three individual tests that test the ability of a PC to perform a series of everyday tasks. The first test edits photos in GIMP and then generates a panorama from six photos, while the second test encodes a high-resolution MPEG-2 video into H.264 using Handbrake. The third test compresses lots of files using 7-Zip while also playing back the output video file from the video encoding test. As the latter two tests are multithreaded, they run faster on multicore CPUs.

Whereas multithreaded games used to be rarer than aviation-inclined pigs, many games can now take advantage of multicore CPUs. To start off, we used the CPU benchmark included with the Crysis demo. This demo plays back a script that sees the main character running amok in an enemy encampment and blowing up buildings with a rocket launcher. The resulting explosions cause lots of debris to fly all over the place. This gives the CPU a hard time, as it tries to work out where the debris will land, and interacts with the environment. Interestingly, this test runs significantly faster on quad-core than dual-core or single-core CPUs; Crytek has clearly done a good job of multithreading the game engine. This test was carried out at 1,280 x 1,024 with all the detail settings configured to medium, apart from world physics, which was set to very high.

RTS games and flight simulators have traditionally been even more CPU-limited than 3D shooters, so we also ran the benchmark built into Supreme Commander at 1,280 x 1,024 with 2x AA. This test plays back a pre-recorded battle in this epic sci-fi-themed RTS, in which several armies battle it out for control of a distant planet. The end to this confrontation is a massive nuclear strike, which brings almost any CPU to its knees when trying to calculate the damage effects of several nuclear warheads simultaneously exploding.

In addition to this staple collection of media encoding and game benchmarks, we also wanted to see how well each CPU architecture performs in a variety of more esoteric applications. First off, we ran the distributed computing medical research program Folding@home (http://folding.stanford.edu) on each CPU. As Stanford has now released an SMP (multithreaded) client for Windows, we ran this version of the client on each of the multicore CPUs, and the standard non-SMP client on the single-core CPUs. The SMP client can support as many as four cores, so we ran one instance on the dual-core, dual processor dual-core and quad-core systems, and two instances on the dual processor quad-core systems. To ensure that there was consistency between the tests, we made sure that the SMP client folded the same work unit, p2653, worth 1,760 points, while the non-SMP clients folded the same work unit, p3405, worth 206 points. We then used FahMon (http://fahmon.net) to measure how many points per day each CPU (or pair of CPUs) would achieve. Don't forget to get your perfect CPU folding when you've bought it, and sign up to the Custom PC team, number 35947.

Meanwhile, for people who are interested in 3D modelling and animation, we ran the raytracing benchmark Cinebench R10, which you can download from www.cinebench.com. Like most professional content creation applications, Cinebench is fully multithreaded, so it runs much faster on multicore CPUs.

The final benchmark was the 1M SuperPi test, which attempts to calculate the value of pi to one million decimal places. SuperPi doesn't tell you how quickly a CPU will run actual applications or games, but we included it in this Labs test, as it's used by a lot of overclockers to compare their systems.

To test the CPUs, we built seven test rigs, one for each platform. All the Socket AM2 CPUs were tested on an nForce590 SLI-powered Asus M2N32-SLI Premium Vista Edition motherboard while a Gigabyte GA-MA790 FX-DQ6 was used for testing the Socket AM2+ processors. The Socket F Opteron chips were tested in an MSI K9ND Speedster motherboard, while the Socket F Athlon 64 FX 7-series CPUs were tested in an Asus L1N64-SLI WS motherboard. All the LGA775 chips were tested on an Intel P35-powered Asus P5K Premium WiFi-AP motherboard, bar the Core 2 Extreme QX9770, which was tested in an Asus P5E3 Premium sporting the Intel X48 chipset. The Xeon 5100-series and 5300-series chips were tested in a SuperMicro X7DAE motherboard, while the Xeon 5400-series chips were tested in a SuperMicro X7DWA-N motherboard. We installed 2GB of Corsair XMS2-8500 in the test rigs, apart from the Opteron and Xeon test rigs, which required registered DDR2 and FB-DIMM RAM respectively. Each test rig also included a 320MB BFG GeForce 8800 GTS and a Samsung SpinPoint P120S hard disk running Windows Vista Ultimate 32-bit with all the latest chipset and graphics drivers.

All the benchmark results are shown on the graphs from p90 onwards, along with a bangs per buck graph showing which CPUs produce the most performance for the money. A further graph shows how many watts of power each system consumed when the CPU(s) were running the small fft stress test built into Orthos. The final graph tallies this power consumption versus the performance of the processor, thus revealing the power per watt of each CPU, or how energy-efficient it is.

Finally, a big thanks to Armari, Boston, Intel, MSI and Scan for lending us the test equipment.