micrOdOts.org / ramblings [of the ABsurd]

I built a new system to replace my aging, hot, and energy-hungry, Dell Dimension 8400 (Intel P4 3.4GHz) as a home server. I went for an AMD Phenom II X3 720 CPU since this seems to be the best value for money currently. I matched it with a Gigabyte GA-MA790XT-UD4P AM3 motherboard and OCZ3P16004GK DDR3 memory.

I like to squeeze out as much performance as possible via overclocking but with the following requirements:

• Stay within the specifications of the hardware.
• Keep power consumption, temperatures, and noise levels low.
• Guarantee 100% stability.

The reason for those requirements is that I run the system 24/7 as a home server, providing services in my local network (e.g. network storage, media streaming, etc.) and on the internet (e.g. web server, this blog, etc.), but I also use it as a desktop when I am around so I want it to be snappy.

First of all, the standard disclaimers apply, e.g. use the following information at your own risk, etc.; I am only responsible for my own computer system. I have collected the necessary information from official AMD spec sheets, internet forums, through the grapevine, and used some common sense. Also, keep in mind that overclocking depends largely on the quality of your components and there may be differences across the same components of different manufacturing batches. So your mileage may and will vary. If you understood all these, keep reading.

Before you begin with overclocking, make sure that your RAM runs stable. Install and run at least one memtest86+ cycle without errors. I am using the OCZ3P16004GK (2x2Gb DDR3) memory which is rated at 1600MHz 7-7-7-24 1.9V. Benchmarks show that to really see a difference between 1333MHz and 1600MHz you need to overclock a lot the CPU NorthBridge (NB), e.g. take it close to 3000MHz. Such a CPU NB overclock hasn't worked out in my setup, so I settled for an 1333MHz speed but with lower latencies. 1333/1600=0.833, this is the factor I multiplied the advertised latencies and set the memory at 6-6-6-20 at 1T. For the rest of the settings, I similarly multiplied the suggested SPD/Auto values. The following settings proved to be memtest86+ stable:

I can probably tighten up the latencies a bit more, but I haven't pushed it. Also, remember that the memory controller in the Phenom II chips runs at 1333MHz stock, so anything above that is considered overclocking. By not overclocking the memory, you can usually achieve easier stable CPU overclocking.

My BIOS says that the default voltage for DDR3 on AM3 is 1.5V. Strangely, it runs the memory at 1.6V by default. My DDR3 is not for AM3 and is rated @ 1.9V. However, it runs stable on at least 1.70V, so I had to give it a +0.1V pump, as we will see later. The memory may sometimes pass memtest86+ with less voltage, but it only passes the mprime's torture test with at least 1.70V.

Next, you need to be able to monitor your system's temperatures and fan speeds. In Linux, set up lm_sensors. The module for K10 is not yet implemented, however, you can use the it8720 instead. As root, run sensors-detect as usual and then edit the /etc/sensors3.conf file. Find the related section and modify to:

chip "it87-*" "it8712-*" "it8716-*" "it8718-*" "it8720-*"
# The following correspondence between it8720 and K10 is speculative.
label in8 "Vbat"
label temp1 "System Temp"
set temp1_min 10
# 60C max is for CPUs with TDP 125W, e.g. AM2+ CPUs
# For AM3 @ TDP 95W, max temp can be even higher up to 73C,
# but keep the maxima here @ 60C to be on the safe side.
set temp1_max 60
label temp2 "Core Temp"
set temp2_min 10
set temp2_max 60
label temp3 "CPU Temp"
set temp3_min 10
set temp3_max 60
label fan1 "CPU Fan"
set fan1_min 100
label fan2 "Front Fan (in)"
set fan2_min 100
label fan3 "Back Fan (out)"
set fan3_min 100
# This is named "Power Fan" in the BIOS and MB.
# However, I connected the top fan
# since my PSU does not provide a head for its fan.
label fan5 "Top Fan (out)"
set fan5_min 100
label in0 "Vcore"
# Stock Vcore is 1.325V.
# AMD specifies safe Vcore voltages for the PII X3 720 at 0.850-1.425.
set in0_min 0.850
set in0_max 1.425
label in1 "DDR3 1.5V"
# Default voltage for DDR3 ram on AM3 is 1.6V.
# My DDR3 is not for AM3 and rated @ 1.9V.
# However, it runs stable on at least 1.70V.
set in1_min 1.70 * 0.95
set in1_max 1.9
label in2 "+3.3V"
set in2_min 3.30 * 0.95
set in2_max 3.30 * 1.05

Run sensors -s as root to set the new labels and limits. With sensors you can see what is going on.

Now, a few keypoints concerning my overclock:

• Benchmarks show that the PII X3 720 has its best performance per watt somewhere in 3400-3600MHz. This is where I want to take it, but stay at or below its maximum spec Vcore of 1.425V.
  
• I do CPU Frequency (FSB) overclocking. If you change the CPU Clock Ratio (multiplier) instead, then the Cool'n'Quiet (CnQ) feature is automatically disabled. (I am not sure if this is how it is supposed to be or it is some "feature" of my BIOS/motherboard.) This is a very good feature to miss, since it slows the cores down, independently and in 4 steps, to even 800MHz (at the stock FSB of 200MHz) when less than full computational power is needed. It also lowers the Vcore voltage accordingly, saving energy and keeping temperatures low. So, with my CPU at 3500MHz, my idle speed would be raised to 1000MHz (= 3500/2800 * 800), and my Vcore varies from 1.125V @ 1000MHz to 1.425V @ 3500MHz.
  
• Benchmarks show that CPU Northbridge overclocking is beneficial. With a little +0.2V pump on the CPU NB VID Control voltage, I managed to get to 2500MHz stably. Note that the CPU NB VID Control voltage should be less or equal to Vcore, so with its stock value of 1.2V and my little pump I reach 1.4V, just below my Vcore of 1.425V.
• Benchmarks show that HT Link Frequency overclocking may actually hurt performance, so keep it at 2000MHz.
• The DDR3 Voltage Control is set specifically for my memory (see earlier comments.)
  

The following BIOS settings (BIOS Version F4G) gave me a stable overclock:

I checked for system stability with mprime and got no errors. After running mprime -t for almost 5 hours, the system temperatures and fans read (check these with sensors):

CPU Fan: 1388 RPM
Front Fan (in): 458 RPM
Back Fan (out):1088 RPM
Top Fan (out): 713 RPM
System Temp: +36.0°C
Core Temp: +47.0°C
CPU Temp: +47.0°C

After 10mins idle:

CPU Fan: 896 RPM
Front Fan (in): 459 RPM
Back Fan (out):1090 RPM
Top Fan (out): 712 RPM
System Temp: +34.0°C
Core Temp: +31.0°C
CPU Temp: +34.0°C

The ambient room temperature was 25C. Only my CPU fan is of variable speed; the rest run always at their fixed highest RPM.

This is how my new server build, charon.microdots.org, looks like:

That's a Cooler Master HAF 932 case (codename "dusty"), a Corsair 850TX PSU, and a Scythe Mugen 2 CPU-cooler. The setup runs fast, cool, and quiet.

The moment of truth. Here are some Phoronix Test Suite benchmarks:

• memory (memory)
• compilation (processor)
• compression (processor)
• cryptography (processor)
• audio-encoding (system)
• video-encoding (system)
• computational (system)
• java (system)

Note that the FSB-overclock of charon does not register with Phoronix Test Suite and it is wrongly reported as a stock 2.8GHz. Also, some tests returned errors during compilation and were skipped. I avoided testing disk and graphics performance as much as possible; those are other stories.

Geekbench results are here; the overclock does not register either.