Overclocking
With the CPU and graphics card sharing a single radiator, we had to prioritize component cooling. Options included high CPU and modest GPU clocks or moderate CPU and GPU clocks. Since games account for only one-third of our final score and many apps are CPU-limited, CPU performance takes priority.
Unlike most overclockers, we test our CPU overclocks for extended periods using the highest possible number of Prime95 threads, set for “Small FFTs.” For this particular system, eight threads took well over an hour to push our coolant temperature to its highest point. We chose 90 degrees Celsius as the CPU temperature limit in an effort to leave at least a little room for extra GPU cooling, and found that the 1.375V BIOS setting got us to 86 degrees Celsius with a 4.30 GHz overclock. Level 2 Load-Line Calibration in BIOS increased the actual core voltage from 1.360V to 1.372V under full load for added stability, as reported by CPU-Z.
Already noted for reaching 8-7-7-18 timings at DDR3-1600, pushing Crucial’s DDR3-1333 CAS 9 to DDR3-1640 forced us to use slower 8-8-8-18 latencies.
The two Radeon HD 5870 graphics processors that comprise Radeon HD 5970 graphics cards use lower clock speeds and voltage levels to prevent overheating under a single air-cooled heat sink. PowerColor changed none of the card’s settings for its liquid-cooled model and the overclocking utility built into ATI’s driver controls doesn’t affect voltage. Fortunately, we’re already extremely familiar with a program that adjusts both voltage and clock speed.
MSI’s Afterburner utility works across multiple brands and even allows custom settings to be launched at Windows startup, with or without the control interface. We had stably pushed the card well beyond the final values seen above before we remembered to recheck all temperatures under combined CPU and GPU loads, only to discover the coolant temperature was too high to keep the CPU below its 100 degrees Celsius absolute limit. Dropping GPU voltage to a very mild 1.187V got our coolant temperatures under control, allowing the CPU and GPU to be fully stressed at the same time, but limiting our card speeds to 980 MHz core and 1,220 MHz RAM.
Note that our discussion of limit temperatures is exactly that: limits at 100% load and not “typical” operating temperatures, which are far lower. Furthermore, these RealTemp readings are taken from directly from the CPU core, and are typically 20° higher than the garbage readings reported by motherboard sensors (this also explains why some users report a thermal reset at "only 80° Celsius" when the CPU actually shuts off at approximately 101°).
Reaching maximum CPU load requires eight Prime95 threads, while FurMark’s GPU stability test requires at least one “free” CPU thread to reach full GPU load. We thus employed seven Prime95 threads and FurMark simultaneously to reach a maximum combined load level that should never be reached, let alone sustained, in daily use.