Chris Thomson built a system around a hot, overclocked Pentium D 830 dual-core processor with a total system power draw of 327W AC and managed to quiet it down to true whisper levels by applying ingenuity and drawing judiciously on the infobanks of SPCR. The great attention to detail makes this one of the best documented DIY articles we’ve posted.
March 26, 2006 by Chris Thomson
(cmthomson at comcast dot net)
| Chris has been a computer system architect at Myrias Research, Amdahl, 3Com, Nokia and others for the last 20+ years. An ex-pat What Chris describes in the following article is the construction of a system based around an Intel Pentium D 830 dual core processor, overclocking the CPU, and finally, making it run as quietly as possible despite the extreme >130W heat. The great attention to detail makes this one of the best documented DIY articles we’ve posted. Thank you, Chris! – Mike Chin, Editor of SPCR |
THE ORIGINAL IDEA
I do a lot of transcoding of audio and video on my home
computer, and got tired of the system being unusably slow for hours at
a stretch during transcoding or other activities such as virus scans.
This was not a particularly slow system (a 2.53 GHz, 1 GB-RAM Pentium 4 HP 754n),
but it sure ground to a halt when busy. I knew
that the solution to this would be a dual-core CPU. When the Intel
Pentium D came out, a guy at work convinced me it would be
“tons of fun” to build a system around one of these chips.
I can safely say that I’ve met the original goal, but it’s
been quite an adventure. Along the way, I discovered the joys of
designing, assembling, overclocking, cooling, quieting and
measuring such a system. As my first-ever do-it-yourself PC, it was a remarkable learning
experience. I started the research in August 2005, and quit tinkering
in November. Then of course as I was finishing this article, I made more changes.
SELECTING A CPU
All I knew at the start of this was that I wanted a dual-core CPU, and had a preference for a Pentium (it’s a long story). Intel had just
introduced a four-member family, the 820 D, 830 D, 840 D and 840 Extreme Edition. All of these are bin selections and bond-outs of the
Smithfield chip, which is basically two Prescotts on a single die, each with a 1MB L2 cache. All variants support 64-bit mode.
I immediately ruled out the 840 EE as a useless marketing gimmick. Four threads instead of two would provide no practical speedup, and the $1000 price tag was just silly.
Scanning the Pentium D reviews on Newegg,
I encountered some scathing comments about the 820 being a different
architecture from the others, and not being well-supported by graphics
processors. Sure enough, perusing Intel’s data
sheet revealed the existence of the PRB (platform requirement bit), which is zero for
the 820 and one for the others. The main difference the PRB makes is in power consumption. The 820 has a TDP (thermal design power) of 95W, whereas the others have a TDP of 130W.
That’s a huge difference, and implies that setting PRB=0 cripples some pretty important performance features. Accordingly, I
chose the 830 D, largely because it is a lot cheaper than the 840D, which has the same chip inside.
NB: If you’re building a system now (in early 2006), you should choose a member of the 65-nm Pentium D 900 Presler family, which has
caches twice as large as the 90-nm Smithfield, and uses less power. You might also consider using a Yonah or Sossaman instead; these are 32-bit-only CPUs
and much more expensive, but they have similar performance while consuming a third as much power.
SELECTING A HEATSINK
Virtually every review of the Pentium D pointed out that it is One Hot Mama.
Most of the reviews complain about the stock Intel cooler being noisy and not particularly effective. I came across many
references to the Scythe Ninja cooler, and a quick Google took me to its SPCR
review. I was hooked: this was the heatsink for me. Unfortunately it was out of stock, so I did the initial system build with the Intel cooler. No
doubt about it, that fan is loud. Also, the temperature was quite high, even with Arctic Silver
5 in place of the obviously compromised stock thermal tape.
Ninja heatsink w/120mm fan, from the SPCR review.
I ordered a 120mm DustProof
AcoustiFan to attach to the Ninja, since I was dubious about dissipating >130W passively, even with the two case fans close by. This turned out to be
a very valid concern. When the Ninja arrived I swapped coolers and was much happier.
SELECTING A MOTHERBOARD
Since this system would be used for general computing and not gaming,
I had no interest in SLI
(the last game I spent much time playing was Adventure;
gosh I’m old!). However, I did want dual-channel memory, which is the PC-world name for interleaving. I also wanted an abundant set of SATA and IDE connectors, DDR2 memory, and PCI-Express support.
The ASUS
P5LD2 met my requirements. All of its components are passively cooled, except of course for the
CPU. Although it wasn’t a consideration at the time, it also has the advantage of using a 4-pin 12V connector. More on this later. This motherboard has
all the overclocking features you’d want: independent FSB, memory and PCI clocks, independently adjustable core, MCH [Editor’s Note: Memory Controller Hub, more commonly referred to as the Northbridge chip] and DRAM voltages, etc.
There are lots of other fine motherboards out there, but this one has worked well for me. The only issues I encountered were excessive motherboard heat buildup and Vcore droop under full load. More on this later.
SELECTING THE MEMORY
As mentioned above, I wanted to use DDR2 memory. With two cores running independent programs, the higher the
memory bandwidth, the better. Also, my old 1GB system swapped more than
I like, so I wanted 2 GB.
DDR2 memories run at very high clock rates, but typically have
high CAS
latency.
Higher latency translates directly into lower system performance, since
the CPU is unable to do anything useful while it’s waiting on the
memory. The main consideration in choosing the memory DIMMs for my
system was a tradeoff between CAS latency and price. Since I
planned to
overclock by boosting the FSB frequency, I also wanted memory that
would run faster than the standard 533 MHz.
There are lots of DDR2 DIMMs that are rated at PC2 5300 (667
MHz) or
PC2 5400 (675 MHz). Most have latencies of 5. For a little more money,
you get a latency of 4. For a lot
more money, you get a latency of 3.
In the end, I chose 2 GB of Corsair
XMS2 TWIN2X1024-5400C4.
These are matched pairs of 512MB DIMMs, rated at 675 MHz with 4-4-4-12
timing at 1.8V. At the time, they cost a little over $200 for the set.
SELECTING THE DISKS
I wanted to have SATA II (3.0 Gb/s) disks, not only for their
high transfer rate, but also for NCQ (Native Command Queueing).
This feature allows the disk to complete requests out of order. This
allows the controller inside the disk to optimize for performance, and
cuts down on random seeks in opposing directions. Because I was
optimizing the system for multiprogram performance, I considered the
NCQ configuration to be better than the more popular bandwidth-optimizing RAID configuration, which disables NCQ. The
biggest beneficiary of this choice is the Windows login function, when a lot of processes are
performing unrelated disk operations.
Second, I wanted nearly silent disks. The reviews of
the SpinPoint family all remarked on how quiet they are.
Finally, when I looked at pricing, the 200GB Samsung
SP2004C leapt out. For some reason, it was cheaper than its competitors, and even its smaller brethren. I bought two of these for less than $100 each.
SELECTING A CASE AND POWER SUPPLY
The last page of the SPCR Ninja review showed it in an Antec
P180 case. After reading the SPCR review
of the P180, I decided that this was the case I wanted to use. It would have the best chance of handling the enormous power consumption of the
Pentium D 830 while remaining quiet.
Size and weight were not a concern. I have a fairly large home office, and the system was going to be on the floor on the other side
of the desk, out of sight and out of the way. I was a bit concerned about cable management, as cautioned by several reviews, but everything
worked out fine.
Scanning the reviews on SPCR for ideas on power supplies, I came across the review
of the Antec Phantom 350 fanless power supply. However, when I went to buy one, all I could find was the newer Phantom 500. The SPCR
review of this supply was also quite upbeat, and I did like the idea of a “backup fan”. It is a snug fit in the P180 case, but with some shoving, everything fits. In
the finished system, the fan never operates, so the supply is silent.
One thing I wanted to be sure of was that the 12V AUX cable would be long enough. It turns out that the 12V cable for this supply has an
8-pin connector daisy-chained to a 4-pin connector. The length to the 4-pin connector is 24″, enough to reach the top corner of the P5LD2
motherboard with comfortable routing. This cable is third from the bottom in this photo.
Phantom 500 cables. The 12V AUX
is 3rd from the bottom, and 24″ to the 4-pin connector.
SELECTING A GRAPHICS CARD
All of my previous PCs had integrated graphics, so I had never
thought about graphics cards. The P5LD2 uses the Intel 945P
MCH north bridge, which does not have an integrated graphics function, so I had
to shop for a card.
A quick survey on Newegg turned up the ASUS
Extreme AX300SE-X/TD,
a PCI-Express x16 card described in the reviews as a “good all-round
graphics card”. After all, something with that many X’s in its name
must be good, right? However, my friend at work was derisive. When I
ran PCMark04 benchmarks, I understood why: it scored about 1300 in a system
that was otherwise well above 5000.
So, back to the search. A little digging turned up the Gigabyte
GV-RX80L256V X800 XL.
Even though it has only one more X in its name, it has much higher
performance, and is passively cooled. Of course, it is also more
expensive.
ASSEMBLING THE SYSTEM
I’ll skip over some of the initial false starts, and describe
the system that I ended up with in October 2005.
I installed the Ninja LGA 775 adapter with the rails vertical;
this is the orientation expected by most LGA775-compatible heatsinks. In my initial build, I
tried to fully seat the fan mounting wires into the heatsink, which
caused them to bump into the heatsink clips, as shown in this photo
(the fins in the lower part of the highlight are part of the north bridge
passive heatsink; the fan was sitting about ½” above it).
As a result, the fan stuck up above the Ninja and some air flow which
could have helped cool the voltage regulators and bridge chips was wasted. More on this later.
Heatsink and fan mount mechanical interference.
Initially, I planned to have the CPU heatsink air
flow towards the back of the case, with a fan mounted on
the front side of the heatsink, blowing through it
towards the back. This worked quite well with the 300SE
graphics card, but when I installed the X800 card, this proved to be a
poor choice, since the CPU was well cooled, but the graphics card
overheated. After some experimentation, I concluded that putting a fan
between the GPU and the CPU heatsinks cooled both quite well. This
works because almost all of the cooling in the X800 is from the
heat-pipe (upper/solder) side of the card, and even minor airflow
provides good cooling. The radiator on the component side of the card provides
at best nominal cooling, since it is poorly coupled to the GPU chip.
This photo shows the same fan as above, from a different
angle, to highlight the positioning between the CPU heat pipes and the GPU heat
pipes. The CPU cooler is the aluminum above; the GPU cooler is the
yellow metal below.
120mm fan between the Ninja (above) and the X800 fanless cooler
(below).
The P180 case GPU cooling duct ended up being a complete dud.
First, the fan housing collided with the heat pipes on the X800 card. Second,
the vent on the back of the case short-circuited the airflow across the
motherboard, and caused it to overheat. In the end, I removed the
entire GPU duct, and blocked the vent with foam.
I also replaced the stock thermal tape on the X800 with Arctic
Silver. This significantly reduced the airflow needed to keep the GPU
cool. This replacement is a bit tricky; the heatsink is actually in
two parts that sandwich around the board. To disassemble, you need to
remove the retaining screws, then pry apart the two-sided tape on the
sides of the base plate that makes contact with the GPU chip. The pry
points are highlighted in this photo.
Where to pry the two
parts of the X800 cooler apart to replace the thermal compound.
Cable management in the P180 case was fairly simple. The only
issues were the some of the power cables. The Phantom 500 has
a standard size main body with a fan assembly added to the back. The cable bundle sticks out
through the fan assembly and needs about another half-inch or so to
splay to various parts of the system. This would not be a problem if
the center fan of the P180 were removed (not a good idea since the
disks in the lower bay would get no air flow), or if the power supply
fan were removed (turning the 500 into a 350 and voiding the warranty), or
the center fan were thinner (which can be done by swapping fans with
the upper bay, or by replacing the stock fan). I went with a thinner
fan. I routed the 12V AUX cable over top of the back fan and had some slack left over.
Attaching the disk drives to the P5LD2 is
straightforward if you read the manual carefully. To support two HDDs with NCQ,
I needed to connect them to the SATA1 and SATA3 connectors. The DVD/CD drive
must be attached to the blue IDE connector, not the red ones. I needed to create a floppy with the ICH7/NCQ drivers from the
motherboard CD in order to install Windows.
One decision during assembly is what to do with the extra
power supply cables. Since they exit the supply at the top, the natural impulse is
to shove them into the gap above the supply. However, this blocks the
airflow across the radiative top heatsink of the supply. It is best to tuck the unused cables either under the supply or
on the sides, where airflow is not important. This photo shows what I
ended up with.
Put unused cables under the Phantom 500 to maximize airflow over its
top heatsink.
There were a few issues with the case that arose during
assembly. The motherboard rear I/O panel is especially cheesy and it buzzed like crazy;
it took some serious bending of the EMI tabs to provide enough
pressure to stop this. The PCI blank panels are tinny and don’t seat
well, and they rattled. This was easily overcome with a bit of tape.
The front plastic panels for the 5 1/2 inch bays buzzed, which was
fixed with a bit of tape. The stock fan for the power supply bay was
too thick. All the P180 stock fans were too loud for my taste even on
the low setting, and in the end I used only AcoustiFan DustProof fans, which produced a much
quieter system. The fat P180 lower bay fan was particularly loud and obnoxious.
Having spent a lot on the fans, I decided to splurge a few dollars more for the AcoustiFan
silicone fan gasket kits. These were easy to install for all but the top fan. A minor case mod,
followed by some shoving and cursing resolved this. The mod consists of bending the mounting tabs slightly, as shown in this photo.
The top case fan mounts with only two screws; bend the fingers to make room for the silicone gasket.
Toward the end of the assembly process, I decided to line most of the box with AcoustiPack
foam, in the hopes of sucking up a few decibels. I couldn’t really tell any difference in the sound, but it looked better. 🙂
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.
QUIETLY COOLING THE SYSTEM
| SIDEBAR: CPU TEMPERATURES The maximum rating consumption and the junction-to-case thermal resistance. To sanity check of all this, I measured the temperature of As a safety feature, the 830 D implements a second All of these temperatures are way |
As described in the assembly section, the Ninja CPU cooler had
three 120mm fans
around it: the top and rear case fans, plus a fan blowing up through
the heatsink. The power supply and disks shared a single 120mm fan. My
next task was to find the best speeds for
these fans to provide good cooling at the lowest noise levels.
Cooling the CPU was the most obvious concern, since it uses
most of
the
power in the system. How much power does this CPU draw? Well, that
appears to be a State Secret. The “thermal design power” is 130W, which
implies that the maximum power is higher than that. Using a PowerAngel,
I measured the total system AC consumption at 262W running two copies of CPUBurn. A review
of the 840 EE states that at 3208 MHz and 1.408V it consumes 178.8W. CPU-Z measured my CPU at 1.22V average, and 3456 MHz. Plugging these values into the “V squared x f ” formula gives 145W. Directly measuring the 12V AUX wires with
a current clamp meter shows 149W with two copies of CPUBurn.
In the most demanding situation, the CPU, graphics card, memory, and voltage converters consume about 200W. That’s a lot of hot air to
try to push quietly through two 4-inch holes. As testing would show, the main challenge would be cooling the motherboard.
Okay, enough theory. Time to talk about fan speeds.
The 120mm AcoustiFans come with a three-speed wiring set that
runs the fan at 12V (black connector, rated at 1500 RPM, 56 CFM and 25 dB),
5V (white connector, rated at 675 RPM, 25 CFM and 10 dB), or an
intermediate voltage connected to 12V through a series resistor (blue
connector). As shipped, the resistor is 56 ohms and the
fan is 92 ohms, so the blue connector provides 7.5V and is rated
at 1050 RPM, 42 CFM and 16 dB. Replacing the inline resistor allows customization of
the fan speed, which varies fairly linearly with the voltage.
I tried a bunch of different configurations and speeds,
including no rear fan, rear fan blowing in instead of out, fast CPU
fan, fast top fan, fast HDD/PS fan, etc, etc. In general, the only
configurations that kept the motherboard from overheating had the rear
fan blowing out, and the top and power supply fans running fastest.
By far the noisiest fan in the system is the hard disk/power supply fan in the lower chamber. This is because the chamber resonates. The typical sound is a
low growl accompanied by a 400-500 Hz tone, depending on the fan speed. Foam doesn’t help. The only time this
fan could be considered quiet is when it is run at 5V (675 RPM). At
this speed, it makes a faint hissing sound, and is quieter than the
disks. Unfortunately, when running that slow, a lot of the heat from the power supply eventually transfers
to the upper chamber, causing the motherboard to overheat and become
unreliable. After many experiments trading off noise levels against fan
speed, I settled on a 78 ohm resistor, which runs the fan at 1090 RPM.
The top fan is the second loudest. When I mounted it with the
silicone gasket, I was surprised that the overall sound increased.
This was because the whole center of the fan started to vibrate, creating a
nearly pure 466-Hz tone at 1015 RPM (obtained with a 103 ohms
resistor). I thought I could damp this by wedging some foam between the hub of the fan and the
center of the case opening. This was only partially effective. Quite by
accident, I discovered that touching the center of the case opening
with something metallic (like a screwdriver) got rid of the ringing.
[Those of you who have ever done EMI testing will recognize the
tape-and-tinfoil nature of this process.] So I built a stack of
quarters held together by paper glue to sit on the top of the case and
damp out the ringing. I am not making this up! Here is a recording
where I alternately add and remove the stack of coins, and here is a photo of the setup.
Top fan damped with silicone gasket, foam, and a stack of coins.
With the case closed, the sound from the fan attached to the CPU heatsink is very well damped. I could run it at up to 1100 RPM without adding a
significant amount of overall noise. These higher speeds resulted in CPU temperatures 2-5°C cooler, but had the counterintuitive effect of raising the motherboard temperature. A series of experiments revealed that running the top case fan faster than the CPU heatsink fan worked best
for overall system cooling.
A side note on the power supply internal fan: when the
Phantom 500 was first
reviewed by SPCR, there was some confusion about when this
fan would activate. My experiments showed that it starts to cycle on
and off when the top heatsink of the power supply reaches
51°C, and to spin continuously when the heatsink reaches
52°C. The only way I could get the power supply that hot was by
turning off the center 120mm fan; running that fan at even the 5V
setting keeps the power supply heatsink below 45°C,
so the internal fan never turns on. This is good, because it is rather
loud, and makes a grating ratcheting kind of racket.
At the conclusion of all these experiments, I had set the
lower chamber fan to 1090 RPM (78 ohms), the top case fan to
1010 RPM (103 ohms), the back case fan to 660 RPM (5V), and
the CPU heatsink fan to 690 RPM (5V). Reported CPU temperatures varied
from 61°C idle to 88°C running two copies of CPUBurn.
The system was highly reliable and very quiet. The only
annoyance was a faint 495-Hz tone that had various nodes and antinodes
in the room; by moving the system around a bit, I was able to position
a node where I normally sit at the desk. Aside from that tone, the sound was a low growl, easily ignored.
MEASURING THE SOUND LEVEL
During the quieting process, I became interested in measuring
and recording the various sounds. One nifty tool I found was this
web site, which makes it easy to tell the frequency of a sound you’re hearing.
To properly record or measure a system this quiet, you really
should have some professional gear, such as that found in MikeC’s
sound lab at SPCR. I wasn’t about to spend the several thousand dollars this
would take, but I thought I’d give it a go with some decent amateur
equipment. For $200, I bought a Radio Shack dynamic microphone
and stand, an M-Audio Audio Buddy preamplifier,
cables, and a CEM DT-805 30-dB sound
meter. This collection was almost, but not quite, up to the task at hand.
My $200 “sound lab”.
The sound meter shows
readings as low as 23 dB, but these are neither repeatable nor
believable. Taking measurements closer to the fans so that readings
were well above 30 dB was really the best I could do, but to get
readings that high I had to be only a few inches away from one of the
fans. Oh well, another tool for the shed. 🙂
AcoustiFan specs the 120mm
fan at 16 and 25 dBA at 1050 and 1500 RPM. The lower chamber fan was running between
these speeds, but also resonated. The overall audible system noise was
less than that of a single fan at 12V/1500RPM, which leads me to guess that
the overall sound level was about 22-24 dBA. Certainly
it couldn’t be heard if there was any noise in the house (furnace
running, TV on, wife at home, cat running around, etc). It was much quieter than the
keyboard, or even mouse clicks.
Here is a recording
made from the desktop, about 4 feet from the system. I clicked the
mouse a few times for purposes of comparison. Note that the microphone
is only rated at -70 dB, and the high-frequency hiss you hear is an
artifact of the microphone/preamplifier/computer; the gain was set to
maximum on both the preamplifier and the computer microphone input. The
low growl/hum component is what I actually heard.
COOLING AND QUIETING, REVISITED
I ran with this setup for about four months, and was fairly
happy with it. However, when I opened up the system to do some
maintenance, I noticed that the CPU and motherboard
temperatures dropped 12°C and 3°C respectively with the side cover off. This
caused me to rethink my approach.
As had long been evident, the hardest thing to do in
this system is cool the motherboard quietly. When I took a closer look
at the board with this in mind, it was clear that the motherboard
heatsinks are designed to catch the downwash from the stock CPU cooler.
This is especially true of the VRM and the MCH, highlighted in this
photo from Newegg:
Hot spots on the P5LD2 motherboard that need airflow: Vcore and MCH heatsinks.
This caused me to question the choice of the Ninja, which has
all its airflow parallel to and well above the motherboard. I reread the
recently updated SPCR
heatsink summary (which coincidentally warns about
overheated VRMs), and saw that the Thermalright
XP-120 rated nearly as high as the Ninja, so I tried one. I was surprised when I
unpacked it by how much smaller the fins are than the Ninja. This
turned out to be significant, because when I mounted its fan blowing
down, the air leaving the fins was so hot it actually increased
the motherboard temperature. I also tried blowing up through the fins,
hoping this would suck air through the motherboard heatsinks, but that
only worked when the fan was run at an unacceptably loud speed. Oh
well, back to the Ninja.
While remounting the Ninja on the motherboard, I noticed that
the fan could be mounted closer to the motherboard and provide some
turbulence around the VRM and northbridge heatsinks if the retaining wires are
stretched around the Ninja mounting clips, like this:
The fan can sit right on top of the northbridge heatsink if the retainers are bent around the clips.
It occurred to me that the temperatures were so much lower
with the side cover off that I might be able to slow down the fans
enough to make the overall system both cooler and quieter.
This turned out to be the case. When I set all four fans to 5V, the
system was cooler and significantly quieter than my prior setup, and was by now much too quiet to measure with the tools at hand. I did have to seal
the gap in the lower chamber between the disk cage and the fan bulkhead so that the HDDs would get some air
flow. A bit of packing tape did the job duct tape would not be removable.
At this point, the back fan was the loudest thing in the case. Its sound was a low thrumming combined with a 400-Hz tone. The
center of the fan was resonating just like the top fan had earlier. I removed the silicone gasket, and it was quieter, but still
resonated. Then I tried cramming some folded tinfoil between the fan hub and the case grill to damp the ringing, but this was only partly effective.
This all-5V setup ran quite well, but over time the motherboard
would overheat, causing erroneous fan sensor readings or other errors. In an attempt to increase the turbulance around the
motherboard heatsinks, I turned up the CPU fan to 985 RPM, but that didn’t help. What worked better was increasing the top case fan to 1010 RPM
while leaving the other three fans at 5V. I ran the system this way for a week or so.
While revisiting how much I could overclock (discussed
below), I continued to encounter motherboard heat issues. Clearly it
was time to take direct action, and mount an additional fan blowing
directly at the motherboard heatsinks. By happy coincidence an
AcoustiFan can be nestled snugly on top of the front pair of DIMMs, which is
where I mounted it. I also moved the CPU fan to the front of the Ninja
so these two fans would blow back across the various heatsinks. I put
a bit of foam on the motherboard fan hub where it bumps against the
graphics card, and secured the fan with a couple of cable ties and some
elastic, as shown in this photo.
Directly cooling the motherboard
with a fan set on top of the DIMMs.
This arrangement not only cooled the motherboard heatsinks, it significantly improved the graphics and PCI card cooling, and let me increase the graphics overclocking. However, it doesn’t cool the northbridge chip optimally because its heatsink fins are perpendicular to the air flow. The geometry of the mounting clips
doesn’t allow for rotation of the heatsink. Rather than do radical surgery, I decided to insert a simple air deflector. This dropped the MCH heatsink temperature a couple of degrees. Here’s a photo:
An old business card used as an air deflector for the MCH heatsink.
The back case fan remained the loudest part of the system even at 5V, and it still had that annoying ringing. So I tried turning it off altogether to see if it was really needed. It wasn’t: the temperatures barely budged with it off. This
resulted in my next configuration: case open, top case fan 5V (695 RPM), CPU fan 5V (685 RPM), power supply/hard disk fan 5V (675 RPM), motherboard fan at 995 RPM, and rear panel case fan off.
TIME TO RESORT TO DUCTING
I continued to be bothered by the lack of cooling for the MCH. The
heatsink was too hot to hold a finger on, which is not a good thing. I
decided to search SPCR for ducting ideas, and came across Edwood’s article about his HTPC. In addition to some fascinating pictures of custom ducts, it also pointed to this tutorial on styrene cutting and cementing.
Since the P5LD2 MCH heatsink has a pair of slots cut horizontally, I figured
I could make a duct that would slide into those slots and divert airflow
from the motherboard fan directly onto the heatsink. The idea was to
have the duct friction-fit into the heatsink and bump against the back DIMMs,
just downwind from the motherboard fan. I actually got it right on the first
try, although it’s a bit ragged. Here’s what the duct looks like from the bottom (the heatsink
side):
The MCH heatsink duct, outside the system (bottom view).
To make room for the duct, I had to shove the CPU fan upward, which
was a good thing since that increased the air flow across the Vcore
VRM. Here’s what the MCH duct looked like in the system, before the
motherboard fan was mounted:
The MCH heatsink duct, inside the system (front view).
This had excellent results. Whereas previously the north bridge heatsink was very hot to the touch, now it was barely warm. The reported
motherboard temperatures also dropped a degree or so.
Encouraged by this success, I started to think about how to vent the
CPU heat directly from the case, separating it from the motherboard
heat. I figured if I could encapsulate most of the Ninja and direct the
hot air out the back, the motherboard cooling would improve. First I
removed the back case fan, which I was no longer running anyway. Next I
built a square duct with flaps to cover the top and bottom of the Ninja
and route the air flow to the back vent. I lined the back portion with
foam, thinking that ought to reduce the CPU fan noise. Here’s what it
looked like before being put in the system:
Ninja duct and its foam lining outside the system (back view).
After installation, I put some foam strips on the case wall
by the back of the Ninja duct to keep it in place and seal off the air
flow. Here’s what the two ducts and fans look like inside the system:
Both ducts and fans installed in the system.
Although in the picture above it looks like the Ninja duct goes all
the way down to the motherboard, it is actually an inch above it,
allowing airflow all around the base of the CPU tower. Also, the
downwash from the MCH duct flows across the motherboard surface.
While designing the Ninja duct, I was hoping I could augment
the airflow through the VRM heatsink with some kind of baffle on the
bottom, but unfortunately the VRM heatsink is so
close to the CPU that there is no room for any kind of deflector. Pity.
With these two ducts in place, I tried running the system with all
the fans at 5V. Unfortunately this didn’t adequately cool the
motherboard at the new clock settings described below; I needed to
return the motherboard fan to 995 RPM (103 ohms).
During these tests I noticed that the entire top of the case was
vibrating and contributing a lot of noise. This seemed to be caused by
the “semi-hard” mounting of the top case fan. I decided to soft-mount
it properly with AcoustiFan gaskets and screws. This required drilling
two holes for the front screws, and bending the inside case tabs
completely out of the way. The fan hub vibrated almost as much as
initially, so I added the foam-and-coin damping described above with
good results. Here is what the soft-mounted fan looks like. The P180
slip-on cover fits over the screws with minimal distortion.
Top case fan soft-mounted with AcoustiFan gasket and screws, damped with foam and coins.
The last step before photographing the finished system was to tidy
up the cables. The AcoustiFan three-speed cable assemblies are quite
unsightly, so I tucked them into an unused disk bay, along with the
PATA cable and power splitters and connectors. I routed the 12V AUX
cable around the top of the Ninja duct. This is what the finished system looks like:
The finished system, with cables tucked mostly out of sight.
This setup is almost silent. When I try to record its sound, all I
get is microphone/amplifier hiss. With the system positioned back
behind the desk, I can hear it only if the house and neighborhood are
totally quiet. It makes a faint low-pitched whirring sound that I
guess might be 20-22 dBA; I have no way to measure it. The reported CPU
temperature varies from 50°C when idle, to 75°C running two copies of CPUBurn.
As mentioned before, the temperatures are way
higher than numbers I see in other people’s reviews. I suspect the
thermal diode circuit in my system is out of whack.
OVERCLOCKING REVISITED
Since the fan rearrangements and ducting improved cooling so much, I decided to see if the clocks could be turned up even higher.
As before, I started with the memory. I was able to run memtest86+
at 683 MHz with the DRAM/CPU/MCH
voltages set to 2.0/1.30/1.60 and the latencies set to 4-4-4-12-4. Note
that I increased the MCH voltage and lowered the DRAM voltage from
before; this eliminated occasional Prime95 failures. Next, I tried
cranking up the FSB. This required increasing the CPU voltage another
notch to 1.325V, to keep the measured Vcore above
1.2V running CPUBurn. With this voltage, the CPU was stable
with the FSB running at 240 MHz. Last, I sped up the graphics card
another 2%.
I ended up with the FSB/DRAM set to 240/640 MHz,
and the GPU/GRAM set to 427/1110 MHz, with no failures or instabilities.
CPU-Z reports the CPU clock as 3.607 GHz.
PERFORMANCE MEASUREMENTS
Here is a table of system parameters, memtest86+, 3DMark05, and PCMark04 benchmark results for the stock system, the
October configuration, and the final configuration.
| Parameter | Factory Settings | October 2005 | March 2006 |
| FSB, DRAM, CPU clocks | 200, 533, 3.00 | 230, 613, 3.45 | 240, 640, 3.60 |
| Graphics GPU, GRAM clocks | 400, 988 | 420, 1098 | 427, 1110 |
| DRAM, CPU, MCH voltages | 1.80, 1.35, 1.50 | 2.10, 1.30, 1.55 | 2.00, 1.325, 1.60 |
| DRAM latencies | 4-5-5-15-4 | 4-4-4-12-4 | 4-4-4-12-4 |
| Memtest86+ L1, L2, memory MB/s | 21054, 18470, 3016 | 24174, 21208, 3463 | 25224, 22129, 3614 |
| 3DMark05 | 4903 | 5292 | 5379 |
| PCMark04 CPU, memory, graphics, disk | 5931, 4752, 6184, 4262 | 6859, 5445, 6690, 4302 | 7139, 5693, 6795, 4441 |
| PCMark04 System | 6060 | 6966 | 7261 |
CONCLUSION
This was quite the project/hobby/obsession. I learned a lot, and exceeded my goals. The resulting system is very fast, almost totally silent, and
finally done (yeah, right!). I look forward to actually using it rather than working
on it. That should be “tons of fun”.
* * *
Editor’s Note: Much thanks to Chris Thomson for sharing his experience with us.
Chris Thomson’s moniker in the SPCR forum is cmthomson, and he may be reached by email at cmthomson at comcast dot net.
* * *
Other Interesting DIY Articles at SPCR:
Doug’s Quiet Wooden Case PC
Quiet A64 X2-3800+ PC for Torrid Thailand
Ducted Zalman 7000CU on A64-3000 w/Countercurrent Flow Cooling
Fanless Heatpipe CPU PC System by FMAH
* * *
