Viewing page 3 of 6 pages.
Previous 1 2 3 4 5 6 Next
OVERCLOCKING THE SYSTEM
With the system assembled, it was now time to see how much performance I could squeeze out of it. I started with fast fan settings
to reduce the number of variables I had to deal with.
The 830 D CPU clock multiplier is locked to 15, so the only
effective way to overclock it is to increase the FSB frequency from its factory setting
of 200 MHz. Most marketing literature quotes this speed as 800 MHz
because the bus is quad-pumped. The memory bus speed is similarly
confusing: the factory setting is 266 MHz, but it is
usually quoted as 533 MHz since it is double-pumped. Here I'll use the
values reported by the BIOS, even though they are
inconsistent: 200 and 533 for the standard settings.
The P5LD2 allows independent setting of the FSB and PCI bus
clocks so you can overclock the CPU without messing up the peripherals. It
also allows the FSB-to-memory clock ratio to be changed. This leads to
three reasonable alternatives to overclocking: speed up the CPU more
than the memory, speed up the memory more than the CPU, or speed up
both by the same amount. The first two approaches are appropriate when
there is a large discrepancy in how much the memory or CPU can be
overclocked, such as when the memory is cheap, or the CPU can't be
cooled. Because I had memory rated at 675 and a Ninja cooler, the
linear speedup approach seemed best.
The first step was to find the clock and voltage settings that would get the most out of the memory. As shipped, the memory is rated at 675 MHz with latency settings of
4-4-4-12-4, at 1.8V; Corsair has since boosted this to 1.9V. The factory BIOS and SPD settings are much more conservative: 533 and 4-5-5-15-4. Using a bootable copy of memtest86+,
I adjusted the BIOS settings to find the highest reliable memory performance in my configuration. I got a lot of single-bit errors running the memory at its rated parameters, but after a bunch of experiments adjusting some voltages, I got memtest86+ to pass at 667 MHz and 4-4-4-12-4 latency. In addition to increasing
the DRAM voltage, I needed to boost the MCH to 1.55V.
Having established that the memory had plenty of headroom, I next started to experiment with overclocking the FSB while keeping the
memory at a 4:3 clock ratio. My main tools in this process were PCMark04, Prime95,
and CPUBurn,
which runs a bit hotter than CPU
Burn-in. I monitored the system with SpeedFan, CPU-Z and ThrottleWatch,
as well as the Probe2 program from the motherboard CD.
Running two copies of CPUBurn (with their CPU affinity set using the Task Manager) produces the highest CPU temperature and lowest Vcore
voltage. The 830 D requires at least 1.2V to operate correctly: with this motherboard, I had to set the nominal Vcore to 1.30V to
consistently keep the voltage reported by CPU-Z at least 1.208V. It fluctuates a lot, since the regulator is apparently controlled on the input side,
not the output side. Note that higher voltages would work, but since CPU power consumption increases with the square of the voltage, there is considerable
incentive to find the lowest reliable voltage.
With the DRAM/CPU/MCH voltages set to 1.95/1.30/1.55, I
was able to turn the FSB up to 240 MHz (3.600 GHz CPU and 640 MHz DRAM)
with reasonable reliability. These settings worked well enough to run
memtest86+, PCMark04, and a few hours of Prime95. However, the ambient
temperature inside the case rose to the point that motherboard
components such as the mouse controller or the USB would fail. Usually
it took several hours or even days for this to happen.
I also overclocked the graphics card. The factory settings for
this card are 400 MHz for the GPU and 988 MHz for the GRAM. Using PCMark04
and 3DMark05
as my tests, I was able to turn these up to 420 and 1098 with the
quality setting at maximum, without any apparent problems. Any higher
settings caused 3DMark05 runs to freeze.
After many experiments, I found that the system would run
indefinitely (48 hours of CPUBurn and Prime95 together, and
weeks on end of World
Community Grid protein folding), with the FSB/DRAM set to 230/613, the DRAM/CPU/MCH voltages set to
2.10/1.30/1.55, and the memory latencies set to 4-4-4-12-4. For the
four months prior to February 2006, I ran with these settings, which clock the CPU at 3.456 GHz.
| Help support this site, buy from one of our affiliate retailers! |
|
