FSB vs CPU?

hitechjb1 said:

...
Adjusting FSB

The ideal situation is that both the FSB and CPU clock frequency are highest possible. FSB should be highest possible so system bandwidth (between CPU and north bridge for XP and P4) for memory and video are highest. Then CPU clock frequency would following FSB x multiplier by adjusting the multiplier. There is also some fine tuning at the end between the exact FSB (down to 1 MHz) and multiplier (down to +- 0.5), so as to tradeoff the final % between the FSB and CPU, since the exact CPU frequency may not be at the exact quantized level given by FSB.

At the beginning, keep the FSB close to its max (then the mulitplier has to be set to low accordingly, since multiplier = CPU_frequency/FSB). This number depends on your motherboard and memory. For recent rev 2 nforce2 mb, and if you have PC3200 CAS2 memory or better, 200 MHz is a good starting point. Then increase mulitplier to locate roughly the CPU max while increase Vcore 25 mV at a time. Run prime95 to check for stability.

Near highest overclocking, mulitplier is kept constant (or increase/decrease multiplier by 0.5 if needed, to allow room for FSB to change). FSB is increased/decreased 1 MHz or small steps (trial and error) until either the FSB or the CPU (or both but not likely) are maxed out (system unstable) for a given Vcore setting. The 3DMark01, 3DMark03 benchmark programs can be used to test the overall system stability, stressing FSB, memory, video in particular.

Sandra is a commonly used program for benchmarking CPU integer and floating point operations raw computing power in DMIPS and FLOPS, memory bandwidth, cache performance, multimedia performance, ... as well as some informations about system hardwares and devices.

For overclocking, don't use default setting. Set them to manual, so you can adust the FSB, mulitplier, memory timing (6-3-3-2 is a good starting point), ...

How to adjust Vcore

After finding the max FSB (or within 5-10 MHz), for CPU overclocking, CPU clock is increased (bigger step at beginning and few MHz step at the end). And when the system becomes unstable, you can increase Vcore by 25 mV stepwise to see whether it can make the CPU to run faster so that the system can become stable again.

Repeat the above until high Vcore is reached to a point that CPU clock frequency cannot go higher practically (say less than 5-10 MHz while temperature rises by 5C per step of Vcore (25 mV)), reaching zone of diminishing return on frequency. From experience, 1.9-1.95 V should be OK for typical Tbred B overclocking (at least for testing), system stability (e.g. running Prime95 for CPU overclocking) is key, while keeping die temperature under control (take 50-55 C loaded as reference rather than the perceived absolute limit). Good HSF is needed to keep temperature low.

There are different level of stability. Minimum is that it can boot in the operating system and run Sandra CPU. Prime95 is considered to be a reasonable way for CPU stability and stress test.

Overclocking is a process, not by sudden boosting the Vcore and hope it will go to 2.5 GHz. Then it would not blow up the CPU. Don't sudden jump Vcore by multiple steps since you will lose sight of the CPU stability and die temperature increases.

During this process, watch for
- system stability,
- rate of increase of clock frequency per step of Vcore increase
(I consider 5-10 MHz per 25 mV Vcore increase as near the overclocking limit)
- die temperature, keep an eye on the rate of temperature increase per step of Vcore increase
(I consider 5 C increase for loaded CPU per 25 mV step of Vcore increase as near the overclocking limit), take 50-55 C as relative rather than perceived absolute limit
- By combining clock frequency and temperature, the rule of thumb is:
10 MHz / C, at full load per 25 mV Vcore increase is about the break-even point for overclocking (point of diminishing return). Below 10 MHz / C is "easy" and economical for overclocking, above 10 MHz / C becomes harder and harder to overclock. And 2-3 MHz / C at full load is the limit.


Example for Vcore increase

Here use a Tbred B 1700+ DLT3C as an exmaple. If you have a CPU that can do 2.2 GHz @ 1.5V, it is a good CPU for oc.
(If you chip can only do 2.1 GHz @ 1.5V, then substract the CPU clock by 100 MHz and 0.5 in the multiplier in the following example, etc, etc.)

Keep FSB = 200 MHz
Raise the multiplier to 11.5, i.e. freq = 2300 MHz
Increase Vcore to 1.6 V+- should get it stable. Run prime95 to test stability.

Then repeat again with FSB = 200 MHz, multiplier 12. I estimate Vcore around 1.7 V +- to get it stable to 200 x 12 = 2400 MHz. Run prime95 to test stability.
etc, etc.

At the early stage, one would expect to get 100+ MHz for each 0.1 V or 100 mV Vcore increase. When close to the overclocking limit, that number will drop to around 50 MHz / 100 mV, and eventually down to 5-10 mV / 25 mV eventually, and the end is almost there. Also at the last % of overclocking, the temperature would increase much faster per step of Vcore increase, I consider 5 C/ 25 mV is about reaching the limit. It will be very costly in terms of voltage, power and temperature to get the CPU to run a % higher.

This is just an estimated scenario, try it out first. If you want to go further, you may need to try a better HSF such as SK7 or SLK-800U/900U with a variable speed high CFM fan such as TT SFII or even a Vantec Tornado (with VR mod).

When you can reach 2.4/2.5 GHz, depends on the HSF and PSU, your are probably 100 MHz away from one of the best oc of a 1700+ on air.

Click to expand...

Gautam said:

What FSB does to 3DMark2001SE scores:

12.5x133.60=1670mhz 10169 3dmarks
12.0x139.20=1670mhz 10121
11.5x145.33=1671mhz 10390
11.0x152.13=1673mhz 10500
10.5x159.29=1673mhz 10601
10.0x167.05=1671mhz 10803
9.5x 175.40= 1666mhz 10915
9.0x 185.03= 1665mhz 11034
8.5x 196.28= 1668mhz 11206
8.0x 208.81= 1670mhz 11364
7.5x 222.73= 1670mhz 11510

In a nutshell, 90mhz on the fsb got me 1341 3dmarks. True, increasing ones fsb will give them a healthy boost, but don't expect anything spectacular, i.e. if you've got radeon 9800 pro and an athlon at 2 ghz with an fsb of 133, and still cant break 10k, theres more to it than the fsb. Very few people can possibly push their fsbs by 90 mhz, let alone 9 mhz. I didn't set out to prove much, and I didn't, but it was fun. I ran only one trial for most. I ran a couple of trials for some tests, but an insignificant difference was seen between the trials, so I just used my initial results. Why 139mhz performs worse than 133mhz is beyond me, and I can say for a fact that it performs worse consistently.

Click to expand...

hitechjb1 said:

Have you run Sandra memory bandwidth or even CPU score on these three:
12.5x133.60=1670mhz
12.0x139.20=1670mhz
11.5x145.33=1671mhz

Maybe the results would show some more hints?

I suppose there were no other parameter changes such as memory timing, ..., among them.

Sometimes there are some discontinuities or roughness in system behavior (reason TBD). I am testing a NF7-S rev 2.0, some how keeping every parameters constant, it is stable for multipliers 9.5, 10.5, 11. But for multiplier 10, it cannot boot, strange or careless on my part ?

Click to expand...

Gautam said:

12.5x133.60=1670mhz 1990 MB/Sec, 6258 MIPS
12.0x139.20=1670mhz 2080 MB/Sec, 6238 MIPS
11.5x145.33=1671mhz 2171 MB/Sec, 6255 MIPS

In other words, exactly what one would expect. But in 3DMark2001, 139mhz did consistently worse. I remember running both 133mhz and 139 about four times each, and having 139 do worse every time. I've never had any issues with any particular multiplier that I know of... but it certainly is not unheard of. Disparities will always exist.

Click to expand...