Overclocking
DISCLAIMER: I am not at fault for anything that happens to you, your computer,
or yourself because of anything in this guide. Your stuff explodes, I am not to
blame. So don’t try any lawsuits.
FAQ
What is overclocking?
Basically, overclocking is allowing your CPU to run faster than the
manufacturer guaranteed it to. There is no physical difference between a 3000+
and a 3800+; both are 90nm
Venice
cores. Same goes with others, the Opteron 165 is the same as the FX-60. Think
about it this way: AMD only essentially makes four different 90nm s939 CPUs:
512kb L2 cache single-core (Venice), 1mb L2 cache single-core (San Diego,
Venus), 512+512kb dual-core (Manchester), and 1+1mb dual-core (Toledo,
Denmark). For example, any two 1mbSC CPUs at the same speed are comparable…
whether it is an Opteron 144 overclocked to 2.8ghz or an FX-57, they should
compare almost equal (if anything the Opteron would be faster, but neglect this
for the moment). A 1mbDC running at 2.5ghz is 20% faster than a 1mbDC at
2.0ghz. So why all the different CPUs? Most of it comes down to marketing
reasons, the FX-57 can be sold for about $700 more than the Opteron 144. The
FX-57 is guaranteed the clock speed of 2.8ghz, while the Opteron is not. This
is the fundamental difference in individual CPUs, some just happen to come out
better than others. If you buy an Athlon processor, you might not have as good
‘luck’ as an Opteron or AthlonFX CPU. The AthlonFX CPUs also have unlocked
multipliers, which are helpful, but not necessary, when overclocking. This is
not a feature that AMD spent extra time putting into the FXs though… rather,
everything BUT the FX series has been ‘locked’, only to make users want to
spend more for their CPUs… avoid the temptation! You are here to overclock!
Let it be very clear there are no guarantees. Your 2.0ghz CPU might not make it
to 2.1ghz. This is extremely unlikely however. Most Athlon processors are
deacent overclockers, Opterons and FXs tend to do better. Another non-guarantee
exists here, too: buying an Opteron or FX gives you a better chance at a higher
overclock, but don’t be flustered if a 3700+ can hit 2.8ghz when your 144
cannot. "Luck of the draw" is unfortunately one of the largest
deciders of overclocking potential.
Is overclocking dangerous?
Contrary to popular belief, overclocking itself can be very safe.
There are two major things you must do when overclocking: raise the speed, and
raise the voltage. Raising the speed (mhz, ghz) of a CPU cannot really damage
it. However, you will probably find that your CPU cannot do X.Xghz unless you
feed it more voltage. While the FX-57 is guaranteed to run happily at 2.8ghz
with only 1.4v, having a 3700+ at such a level may be wishful thinking (but
again, this is not always the case!). Voltage increases are generally the
reason a CPU gets destroyed by overclocking. Do not fear though, voltage
options are clearly labelled in your motherboard setup. This is not something
you can accidentally mess up… if you keep your CPU under 105% of what the core
is rated at, you should be more than safe… but we will get to closer
approximations later. Remember that the CPUs were designed to handle voltage
fluctuations: not even the most expensive power supplies can keep a CPU at
exactly 1.4v.
If you do want to raise voltage, adequate cooling is a must. Whether the stock
heatsink, a replacement heatsink, water cooling, or exotic cooling is necessary
depends on what you are wanting to do. Yet again, more later.
Overvolted CPUs do tend to not last as long as stock-voltage CPUs. But when we
are looking at about a 15yr lifespan for most, you might not mind that number
being cut down to 10yrs, or even 7yrs… think about a computer that is even
5yrs old, and how much you would mind if its now-$30 CPU crapped out tomorrow.
Will overclocking void my warranty?
Yes.
Getting started
Part Selection:
CPU:
The best performance-for-money overclockers at the time of writing this article
would have to be the s939 Opterons. They ship with a very low stock voltage,
despite the fact that their identical Athlon brothers need more voltage to run
at the same speeds. This gives you extra headroom, because while 1.4v is
technically an overvolt for an Opteron, it is a safe voltage endorsed by AMD in
their Athlon line! So the amazing overclockability of Opterons on stock
voltage, combined with the .1v ‘boost’, gives spectacular results.
If you have extreme loads of cash and do not mind spending a few hundred on
something that will be antiquated in a year, the FX line might be for you. They
offer unlocked multipliers which offer more freedom in overclocking, and they
have very high speeds to begin with. I would however not suggest an older FX
CPU though, such as the FX-53, as for much less money you can have a 4000+.
Surely in time, the FX-55 will have a non-FX counterpart, etc. Pretty much, the
only FX worth buying (if an FX is worth buying) is the current fastest one of
either single or dual core.
Mainboard:
This is as important as the CPU. A good board is what provides you with all the
options for overclocking. Due to the extremely varied and next to unpredictable
nature of motherboard production, even a brand is hard to recommend… but most
enthusiasts prefer DFI. Jetway is another manufacturer that is beginning to get
into the high-overclock market for a very good price. ASUS makes
deacently-overclockable boards with simpler (read: less) options than the
previous two mentioned. Make sure the mainboard has LOCKED
SATA/IDE/PCI/PCIe/AGP ports. Basically, unlocked of these will cause huge problems,
including corrupted data. Most board have locked SATA ports, for instance… on
my DFI, 3,4,7,8 are the locked ports. USE THESE.
RAM:
Many think that to overclock, you need superultraamazing RAM. This is not the
case. Thanks to memory dividers, even regular DDR400 memory can be used in
combination with a highly overclocked processor. However, to give more freedom,
and to make sure that it works well in a premium mainboard, you might want to
invest in some good memory. OCZ and Gskill both make amazing products; Patriot
does so for a lesser price. It is at least as important to look at what kind of
memory you are buying though… Samsung TCCD allows high clocks at low
voltages, but do not have great timings (another factor in overall speed). BH-5
memory has great timings, but does not acheive as high clocks, and requires
higher voltage. Huge debates go on and on about which is better, but the truth
is, both tend to perform about the same when doing their best. (It should be
noted that the author of this article currently prefers TCCD, as its lower
voltage is easier on the CPU and motherboard, and its great variability in
overall mhz allows freedom in overclocking… but this is only an opinion.)
Power supply:
Get a good one. DO NOT SKIMP ON THE POWER SUPPLY. Yes, it’s about the only
thing that doesn’t improve performance when you pay more for one, but a quality
power supply means a lot… whether it is higher overclocks, a more quiet and
stable system, or even a working system at all! A junky power supply could be
the demise of your new $2000 tower, or the reason you can’t get to X.Xghz.
Stable voltages and adequate power are musts, especially for overclocking. When
selecting a power supply, look at the amperage per volt, not the overall power
rating! For instance, a good ~500w power supply will probably have at least
30amps on the +12v ‘rail’. A cheapo 600w, though, might only have 15amps. A
quality manufacturer is also a really good idea, like Epower/Tagan and OCZ
(which are very similar to Epower/Tagan). Read reviews, ask people. It’s
generally a difficult thing to decide by numbers alone, so you’ll have to
research this one.
Cooling:
For a small, or even moderate overclock, the stock heatsink is okay. For a
higher overclock, or just to enjoy cooler and silent operation, a better
heatsink is advised. For extreme, record-breaking overclocks, watercooling and
phase-change cooling are preferred, but are costly and generally
high-maintenance. If you aren’t even sure what these do, chances are you want
to just stick with air cooling for now.
Understand that increasing the speed of a CPU barely increases heat, and
doesn’t demand the need for better cooling. Increasing voltage, however,
greatly adds to heat.
BIOS
Whenever you get your computer put together with XP installed etc, during a
startup, push the key to enter the BIOS (usually [
DEL
]). This will bring you to a menu of the
computer’s low-level operating conditions. Without going into too much detail:
set up your boot order, disable things you don’t want or need, save and
restart. Enter BIOS again, and begin exploring. There is generally an
overclocking section/page, almost always named something different depending on
the manufacturer. DFI boards call it "Genie BIOS", MSI calls it the
"Cell Menu"… either way, find it. The following are a few options
that we will be using for the remainder of the guide:
FSB/CPU clock/HTT speed (hereon referred to as FSB): This, times the
multiplier, derives the overall speed of your CPU. For example, 250 FSB x 11
multiplier = 2750mhz. This also controlls the memory speed… 200mhz here
equates to DDR400 memory. 250mhz, DDR500. Mhz*2 = DDR rating. Also, we can use
dividers to make your memory run slower if your CPU can be let loose…
discussed in detail momentarily.
Multiplier: As shown earlier, affects the overall speed of the CPU.
HTT Ratio: FSB x HTT ratio x 2 = HyperTransport speed. For the latest CPUs,
this is 2000mhz. Notice stock, it is 200mhz x 5 x 2 = 2000mhz. If your FSB is
at 300mhz, then you need to kick this down to 3 to stay under the 2000 limit.
For FSB between >200 and 250, use 4x. 250 to 333, use 3x. You can figure it
out from here, if you get that far that is.
CPU Voltage: The voltage sent to the CPU. Take care in not making this too
high. Look up what is stock for your CPU, and make an informed decision about
how high you want it… but don’t set it there yet. Let’s wait and see how hot
your CPU gets with its current cooling before you go crazy. On some boards the
voltage is a combination of a base and a percentage, such that 1.400v and 110%
would give 1.540v.
RAM Voltage: Voltage to RAM. Take same precautions as for CPU voltage. If you
have hot-running RAM and/or >2.8v to it, use active cooling (ie, a fan).
Chipset voltage, and LDT Voltage: These rarely need to be increased. By the
time you are ready to do so, you will likely not be reading this guide.
RAM Settings
Memory Divider: The easy way to use this for now would be to multiply the ratio
times the FSB to get how fast your memory is running. For example, a (5/6)
divider will make your RAM run at 5/6 the FSB. It may be represented in terms
of DDR400; in which case, (5/6) would be called DDR333, or a 166 divider. It
should become pretty clear by the time you see the list though.
Memory Timings: Depending on your board, you should see anywhere from four to
fourty of these. I won’t get into detail here, but when you bought the RAM, its
timings may have been advertised as 2.5-3-3-7. This is in order of
CAS-tRCD-tRP-tRAS. It might be a good idea to either leave these at defaults,
or set them to what is specified.
Command Rate: This is 1T or 2T. 1T is faster than 2T. 2T provides greater
stability at a loss of performance. Most RAM is made to run at 1T. Try to keep
it at 1T.
Finding your CPU’s maximum
Set the memory divider to something low, like (1/2) or (2/3). Make the command
rate 2T. You are now going to see how fast your CPU can run independently of
the memory.
Starting at 200xDefault, raise the FSB maybe 5mhz. Change the HTT Ratio to 4x
to avoid exceeding 2000mhz on the HyperTransport. Save BIOS settings, and
restart.
If you start booting to Windows (or OS of choice) this is a good sign… if
your multiplier is say 10x, you have just acheived a 50mhz overclock! Refer to
‘Stress Testing’ section right now. An overclocked CPU that isn’t stable is no
good. If it is ’stable enough’, then restart, and raise your FSB a bit more.
Eventually, you might not be able to boot Windows, or pass a stability test. It
is at this time you might want to consider increasing the voltage. REMEMBER THE
WARNINGS DISCUSSED ABOVE. If your CPU is stock 1.4v, maybe try 1.45v. It should
help if you’ve been going slow. If you’re impatient and went for 250xDefault
right off the bat, then judging a proper voltage is going to be difficult. For
beginners, keep the voltage within .1v of stock (or for Opterons, up to 1.5v).
WATCH THE HEAT! If your CPU was getting near 50*C running the stress test, you
might not want to increase voltage. In fact, as a general rule, your CPU should
never reach 50*C, even overclocked and overvolted. This isn’t near a dangerous
temperature, but it is a good way to make sure you don’t kill your CPU slowly
over time.
So after a little while, you will see how far your CPU will go with a maximum
of 1.5v, or whatever you have chosen for yourself, or whatever keeps your
temperatures under 50*C. For the example, let’s say your CPU’s max is 260×10,
or 2600mhz.
Finding your RAM’s maximum
Set the CPU multiplier to at least one less than default, but no lower than 7x.
We have to make sure that in this situation, FSBxDivider does not exceed even
stock ratings. Also, go ahead and change CPU voltage back to default. Set the
memory ratio to 1/1, or 200, or DDR400.
Now again, slowly raise the FSB. If you get a successful boot and stability
(see ‘Stress Testing’), go for more! DDR500 RAM should be able to hit 250mhz,
you can start there if you want, etc for other memory speeds. If it is
unstable, you may choose to add voltage, and/or ‘loosen’ (increase = slower)
timings. There are so many choices and complications when doing this I suggest
you seek an outside source, other than higher numbers providing more stability
and higher clocks. If your RAM is say rated to run at 2.5-3-3-7, then maybe
increasing it to 3-4-4-8 to acheive a 50% overclock on your RAM might not be a
bad idea. Play around with it. Go for the highest mhz you can, but don’t send
your timings to trash just to gain an extra 2mhz.
When you have found the maximum speed your RAM will run at, continue to the
next step. For the example, we will say the RAM runs at 240mhz at most.
Balancing CPU/RAM speeds
For our example, the CPU hit 2600mhz with 260×10, and the memory hit 240mhz. So
how do we do both, instead of just one at a time? Well, remember that the FSB
controls both. For the CPU’s sake, we will set the CPU to 260×10 (remember to
restore the necessary voltage, to both CPU and memory!) and then go back to our
RAM settings section. Think about the divider options you have: in our example,
our memory cannot run at 260mhz, but only 240mhz. The closest we can come would
be by the use of a (9/10) divider. That would make our memory run at about
(9/10)*FSB, which is about 234mhz. Kind of sucks that we lose a little, but
rarely will you optimize this balance. Go for CPU mhz first, unless we are
talking about taking 50mhz off your RAM speed just for 10mhz on your CPU (I am
still researching the best tradeoff ratio, I’ll edit the guide when I find it).
If your RAM maxes out at say 260mhz, then you can run it at (1/1). If your RAM
exceeds the FSB that your CPU did, you can use one of two options (or both): a
positive divider, such as (5/4) found on some DFI boards, which would make a
260mhz FSB get the RAM running at about 325mhz (wow!). Another option would be
to lower the CPU multiplier and raising the FSB to give you approximately the
same overall speed. Using our example, 289mhz FSB with a 9x multiplier would
still get you 2600mhz on the CPU… but now we can run the memory (1/1) with
the FSB, keeping the memory at a high 289mhz.
Now that you understand the concepts, time to get the exact RAM speed in
relation to the FSB. This was excluded before due to its complication and
unnecessarily-preciseness. The formula is as follows: …… Chances are it
will not make a huge difference in your RAM speed, but it’s good to know.
Stress Testing / Heat Output
An overclock that is not stable is worthless. You cannot claim you paid $300
for a virtual FX-60 when it barely boots up. To make sure your computer will
not be restarting on the hour crashing games, we have to test it to make sure
all is well. I’ll outline a few tools and purposes:
WHEN TESTING CPU SPEED:
Stress Prime 2004 - A good program, but it takes a long time to guarantee
stability. This isn’t necessarily a bad thing though. It uses prime number
calculation to make sure your CPU is not miscalculating numbers. Use the ‘Large
FFT’ test. If using a dual-core CPU, install two instances, to two different
folders. Start both. Uncheck the ‘No Affinity’ box for each. Set one to CPU0, the
other to CPU1. Still using the ‘Large FFT’ test, run both at the same time.
Chances are, it will find errors when you might not even see anything wrong
with the computer. This is normal. You might be able to play Doom on something
that Prime says isn’t stable, but you won’t be happy when your computer
restarts when you’re writing a superlong guide on overclocking that you never
saved (not based off a true story, fortunately). A ’stable enough’ run is
letting it run for 30mins or so, when you decide to restart and go for a
slightly higher overclock. When you think you have found your maximum, let it
run for at least 8hrs before you call it stable.
WHEN TESTING RAM SPEED:
MemTest86+ - This is a bootable program, not to be run in Windows. If you have
a DFI board, it is built-in, and enable(able?) in BIOS. Let it run for at least
one pass until you overclock the memory further, and let it run for about 8hrs
when you think you have found your maximum. This is the same deal as when
testing the CPU… don’t fuss when it says it is instable and you disagree, or
you’ll be sorry later.
WHEN TESTING OVERALL:
Stress Prime 2004 - Keep using the ‘Large FFT’ test. Doesn’t hit your RAM that
hard, but you’ll find that maybe a maxed out CPU and maxed out RAM are suddenly
unstable.
Memtest - Same idea as above, you want to make sure that the RAM doesn’t start
acting up once your restored your CPU max.
OCCT - Another great program, that stresses CPU and RAM. It tends to find
errors much faster than Stress Prime 2004. It’s greatest flaw is that it is
useless for dual-core CPUs. If you have a solo, this is the best test for final
stability. The 30min test will give you a good indication of where you stand,
but let it run for a few hours if you want to be sure you are stable.
ALWAYS CHECK TEMPERATURES!
Heat kills a CPU. Remember to occasionally check temperatures in BIOS when you
just set voltage a bit higher. When you are doing any Windows-based testing,
check out some good software solutions. I prefer MBM5. It can be set up to read
a variety of temperatures. Remember to keep the maximum temperature under 50*C,
and idle temperature under 40*C. Be wary though! Software is not a very
accurate way to measure temperatures, or much for that matter. If it seems
unreasonably low (like 20*C idle with the stock cooler, or not breaking 30*C
under full load) then something is wrong, and you’ve basically lost any chance
of software helping you out. For reference, single-core 90nm CPUs tend to have
an idle of around 30*C, and a full load of 40*C at stock settings with the
stock cooler. Dualies, 30*C idle at up to 50*C during dual SP2004s. Always use
the temperature of a 100% loaded CPU when running SP2004 for your maximum
temperature.
Hardware monitoring via probes is the most accurate way, but is however
fundamentally flawed because of placement options. The hot part of a CPU is
covered by a ~1mm thick slab of metal that messes everything up, the IHS.
Putting a probe on this, especially on the sides, will give you a reading not
of the CPU core temperature, but something much cooler. Do not put the probe
between the heatsink and CPU. Just about the only way to get an accurate
reading with probes would be to remove the IHS and put the probe on the side of
the core. I do not suggest this at all. It is extremely risky. At the very
least, remove your IHS for better temperatures, and consider good probe
placement a bonus.
