FSB overclocking and multiplier manipulation:

FSB overclocking is the easiest and most common way to overclock. However, if you increase the FSB, most components are overclocked. That can be a problem for numerous reasons on an AMD motherboard. All AMD boards but the latest N-force 2 chipset motherboards do not have an AGP/PCI frequency lock. That means that the AGP/PCI devices will get their bus overclocked as well. That offers nearly no increase in performance, however on the other hand video cards and hard disks among other components do not like that at all. The hard disks might get corrupted or permanently damaged, the video cards unstable, producing corrupted images or not work at all. That is why overclocking with a motherboard that doesn’t feature these locks is not recommended at all, the risk for the other components is too high, and so if you proceed make sure you can take the risk. Another reason is that for optimum performance on AMD motherboards the memory must be kept at synchronous operation with the CPU bus, which is the FSB itself. So if you wish/try to raise the FSB to 200, you’ll need at least Pc3200 (DDR400) memory to be safe about that, else the memory might not handle the overclock.
The multipliers on most AMD CPUs are locked at a fixed factory setting, however there is a way around that. Many recent motherboards will unlock the newest processor cores themselves, up to 12.5x which is plenty for overclockers, since most will want to reach a high FSB speed. You can also unlock the older Thunderbird and Palomino core processors, however even if you do the performance gain that you’ll get doesn’t really worth it anymore. Too little, too late.

Memory overclocking and performance:

As I’ve already mentioned, for optimal performance on any AMD motherboard the memory must be running at the same speed with the FSB speed. That is often referred as synchronous mode or 100% memory speed setting. Many boards offer dividers so the memory can work lower or higher from that speed. If you use them, even set higher the memory will lose performance instead of gaining any. The memory cannot run any faster than the FSB, even set higher it will be bottlenecked to that and lose performance due to the increased latencies.

For example, an AMD Barton 2500+ processor running at the factory speed has an FSB of 166. Since the latest AMD processors are all using DDR bus, that’s doubled to 333MHz. The memory to match it is rated for Pc2700 (DDR333) operation. Even if you get Pc3200 (DDR400) memory and use the 120% divider to run it at that speed, its performance will be the same (if not a little lower) of the Pc2700 memory since the processor cannot receive or send data at a higher speed. Of course if you get the Pc3200 memory and overclock the processor FSB to 200MHz, the full memory bandwidth will be used and its performance will be increased dramatically.

What about Dual Channel DDR on the N-Force 2 motherboards? Well, in theory DCDDR is meant to double the memory bandwidth by splitting the data sent to the memory to two modules. Of course that requires two modules of memory to work. In real world usage that is a 70%-90% performance gain. However, since the memory cannot run any faster than the CPU FSB, for AMD platforms it is nearly useless. For example a processor running at 166MHz FSB will require DDR333 memory, as mentioned. When DCDDR comes into play, 2 sticks of DDR333 memory will give you about 333MHz (DDR667) speed. That will simply be bottlenecked down to DDR333 speed again. As many people will think, this way you could use cheap memory that is much slower than the processor, since the speed is doubled. Like for this 166MHz FSB processor, 2 sticks of Pc1600 (DDR200) which will give about DDR400 performance should be plenty and even very slow will still be underclocked to meet the processor FSB by the use of the 50% divider. Well, that would be true if Nvidia would let their N-Force 2 chipsets to leave DCDDR support enabled if the memory is using any divider. In simple words, if you won’t run the memory synchronous with the CPU, DCDDR is disabled and the memory performs as any normal memory modules. Final words, DCDDR on AMD motherboards does help but very little, it’s only meant to help when an on-board video card is present since that shares the memory bandwidth with the CPU and will fill the gap, however still on-board video cards are not suitable for gaming applications, they can’t offer the performance required and are far inferior than their real video cards counterparts.

Memory timings are somewhat a pain to be tweaked to perfection. They entirely depend on the memory module. However, on AMD motherboards lower timings help a lot, since the bandwidth is limited by the CPU. It’s better to get memory that can stand low latencies at a respectable speed than to get memory which can reach crazy speeds but won’t take low latencies. You can never get the FSB too high on an AMD platform unless you are a veteran overclocker. So if you want every little bit of performance out of your AMD system, get low latency rated modules, 2 of them for usage in DCDDR mode if you own a N-Force 2 motherboard, it will help a little but it will still help. (Note: On N-Force 2 motherboards there is a ‘bug’. When DCDDR is enabled, on many motherboards by raising the Active to Precharge delay instead of lowering it will increase performance. If that will happen with your motherboard and memory modules, you have to find out yourself by testing it set to 5-6 which is a normal setting and then to 8-11 which are very high settings).

Voltage settings and cooling:

To overclock any component, after a certain limit you may have to increase the voltage supplied to it. For AMD CPUs, that’s not much of a hazard as long as the cooling of the CPU is sufficient, since those CPUs are not very vulnerable to EMI. As long as cooling is sufficient, which means that the CPU temperature -under load- should not exceed the 55c wall, you may increase the Vcore up to 0.3v higher than the stock with absolute safety. Anything over that is considered dangerous, but won’t kill your CPU instantly unless extreme settings are used (2.1v and over) which even with great cooling they will shorten the life of the CPU severely. However at all times increase the voltage at steps and not all at once. You can’t always know if your cooling is appropriate to hold the voltage you are aiming for, so you have to keep an eye on the temperatures. For cooling you may want to consider a good (hence large) copper heatsink, since AMD cores heat up very fast and the heatsink material has to be the best heat conductor available. Good isn’t necessarily loud as well, there are enough great performing heatsinks that use quite silent fans.
For memory voltage, anything up to 2.7v is covered by any company warranty. Some companies cover more, like Geil which for some modules they give warranties up to the very high 3.1v. I’d suggest setting the memory voltage to 2.8v unless you use generic memory. This kind of voltage will lessen the life of the module to about 3-5 years if it’s of decent build quality. I don’t think any user would keep a system that long, especially overclockers which will buy new ram within 6-12 months (I told you that it is addictive). By using that voltage setting, you can limit your modules abilities and still be on the safe side. Of course if any company warranties above 2.8v for the modules that you have, like some OCZ or Geil modules, use up to that limit if you motherboard will allow you to do that. But if you have generic memory, skip that and leave voltage at the stock 2.5v.
If your motherboard has a VDD setting, you better improve the northbridge cooling if you wish to use it. It will help a lot to achieve high FSB speeds. For a northbridge cooler, you may get something like the Titan Copper GPU cooler (among many other good chipset coolers) or even use a ‘small’ CPU heatsink. CPU heatsinks are somewhat to extremely large though, so you need to find a 60mm heatsink/fan or smaller. Even a ‘small’ 60mm heatsink is far superior to most chipset coolers because of its massive size difference. Older pentium 3 coolers are small like that, but you may even find more recent coolers around that are small enough. To mount it, you need to use something adhesive to glue it on, like Artic Alumina epoxy or Zalman adhesive (that’s quite permanent of course) or thermal 2 side sticky tape (which will come off with no real effort, so make sure the heatsink is light!). Alternatively, you can drill holes on the heatsink base directly where the motherboard holes exist and use screws to bolt it on the motherboard (with a rubber ring under their head to prevent it to touch the motherboard). I recommend the thermal tape since it will come off with no trouble or a sign left on the chip and does the job well enough. A safe VDD setting is any up to 1.8v when good cooling is applied.

Basic procedure and testing loop:

To find out the limit of your system, it is not very hard but it needs time. I would like to add that I would not overclock with any motherboard that doesn’t feature an AGP/PCI lock because it becomes hazardous to the other components of your system. Also to do some effective overclocking you’ll need a recent motherboard and CPU, so that it will be or become unlocked. When you have the whole system and its cooling equipment ready as well as the voltages are set right, then you are ready to proceed.
To start, you should find out your FSB limit. Set the memory timings to SPD or Normal and the multiplier very low, like 7x. Then keep raising the FSB at 5MHz steps at the time and testing the system stability. To test the system stability for the memory/FSB you should run a SuperPi 2mb benchmark and 4-5 loops of 3DMark 2001. If you find that the system crashed, was unstable, did not even boot or did not even post, it just couldn’t handle the FSB obviously. The last stable frequency is about your maximum FSB limit.

Then you should try to find out the maximum MHz your CPU can give. To do that, set the multiplier as high as possible. That will be 12.5x for any CPU that uses a multiplier than that or the stock multiplier of the CPU if that is higher than 12.5x (15x is the stock multiplier of the 2400+ for example). Set the FSB at a corresponding speed so the outcome will be the CPU stock speed to the lowest of 133MHz. For example, a 3200+ has a stock multiplier of 11x and 200MHz FSB, giving a 2200MHz output. You should set it to 12.5x and 175MHz FSB, so that it will again give about 2200MHz output. But for a 2400+ with a 15x multiplier and 133MHz FSB, its stock speed is the lowest limit already. Then raise the FSB 3 at the time, to find out what the maximum speed that your processor can achieve is. Always have a look at your temperatures! When the processor speed is raised, heat output is raised as well! Since you already should have raised the voltage by 0.1v-0.3v, you should make sure it doesn’t exceed the acceptable levels. To test the stability of the CPU, you should run a Prime95 test for 15-25 minutes. If it gives no errors, go ahead. When it gives errors, the last stable frequency should be the processor MHz limit.
Last, using the conjunction of the maximum stable FSB and processor MHz limit you should use a multiplier that corresponds in both of those. For example, if you find out that the maximum stable FSB is 210 and the maximum processor speed is 2420MHz, you should set the multiplier to 11.5x which with 210FSB will give you 2415MHz speed. Again, to test the system stability thoroughly this time, have it run a Prime95 test for some -hours- and 30-40 loops of 3dMark 2001. It should be stable, however if it isn’t just lower the FSB by a couple of MHz and try again. And keep an eye on the temperatures always, things heat up to their maximum through this testing.

After all that, you can go and tweak the RAM timings if you wish. Set them to Expert/Manual which will allow you to manipulate them yourself. Lower one of the 4 settings each time and let windows boot (which is not exactly guaranteed to happen if the RAM can’t handle it), then run a 2mb SuperPi to check the stability. If it gives no errors, good, go and lower another setting as well and try again. If the system becomes unstable or won’t even post, the memory simply cannot work with the current settings. Get the timing that made it unstable back up and get another down.

Now, after all these you should have lost one or two days of your life, if not more. To speed things up, you could ask a veteran overclocker or two about the overclocking ability of your components, so you can start the testing procedures from higher than going to small steps at the time. When I say veteran I certainly do not mean a friend of yours that did overclock a couple of processors a little bit, but a person that has vast knowledge of the overclock abilities that components can have and has been doing it for some time. For example, even the crappiest 2400+ processor that I’ve seen, with 1.85v will at least hit 2200MHz as long as the cooling is sufficient. Instead of starting at 2000MHz with one of those, you can start at about 2200MHz if you can supply enough voltage to them. They may go well over 2400MHz or stay close to 2200MHz, depends on too many things such as the system abilities and cooling.