lower latency and lower fsb, or higher latency and higher fsb?

[irabufan]

what is a better setting? 166mhz @ 2-2-2-6 or 183mhz @ 3-3-3-8, assuming the cpu runs at the same clock speed?

 

[MDE]

Test it for yourself and tell us .

 

[irabufan]

i get better memory scores in sandra when i run it at 166mhz @ 2-2-2-6
but not sure about other benches

 

[orion7144]

If you use the search button there are already several threads that go into detail.

 

[dunkster]

The benchmark to determine which set of FSB/timings is best overall is one that measures total system bandwidth - not just memory bandwidth. 3DMark2001 is a good choice.

Hope this helps!

 

[DAPUNISHER]

Personally, I wouldn't rely on any synthetic benchies to determine the optimal config for your setup. Using FRAPS while gaming if that is your primary concern for performance and comparing the logs is a far better solution than 3Dmock or SiSuck. If you are more concerned with performance for another task/tasks list which and one of us will suggest the best way to evaluate the difference.

 

[Soulkeeper]

i would definately use the lower timings in that case
the gain from 2-2-2-6 over 3-3-3-8 should be greater than a 20mhz increase
but yeah benchmarks are always good

 

[pallmall]

Look at this table :

multi, 166, 183(multi-1.0), cpu diff
10.0, 1660, 1647.0, 13
10.5, 1743, 1738.5, 4.5
11.0, 1826, 1830.0, -4
11.5, 1909, 1921.5, -12.5

12.5, 2075, 2104.5, -29.5

If you are not considering the multiplier, the cpu clock will vary significantly if the multi goes beyond the table.

So the closest cpuclocks will look like below for an example.
multi, 167, 183(multi-1.0), cpu diff
11.5, 1920.5, 1921.5, -1

The number one factor for the realtime performance is the cpu speed.
So the performance will vary depending on the multiplier you choose, which will be mostly affected by the cpu overclockability.
This problem is caused by the current pc architecture, which means we need the pc setting the clock directly.
It is because the mem clock is designed to be tied with cpu fsb.
If we can set the mem clock totally independant and cpuclock directly then we will be free from the multi*fsb effect(*check Edit).

Current advanced solutions for memory clock independance are not still all perfect.
Why are they setting the gaps by adopting the cpu:mem clock ratio?
Some bioses are adopting mem/cpu %, which is still similar to the previous one.
Only a few boards are providing real solution by allowing setting the mem clock by 1mhz increments.
But they still have the cpu fsb * multi architecture(*check Edit) unresolved.

The other problem for comparing the performance is from the memory timings nature.
166, 2.0,2,2,5 vs 183,3.0,3,3,8 ?
Are they both 100% optimized?
Because of the settings not being continuous(analog), the increment of the last 4 numbers results in big differences.
So they are just rounded for the memory clock and hardly can be all optimized together.

Because of these limitations, we need to be careful when comparing or analyzing the performance.IMHO

[Edit for cpu fsb * multi architecture]
By further research, I've got the conclusion that under the current architecture the CPU FSB is unavoidable.
It needs to be fast enough to cover the traffics between the cpu and northbridge(and southbridge).
The higher will be better for the overal performance.

 

[pallmall]

[addition for the above Edit]
My suggestion for the board maker for easy and fine tuning is to let the user give cpu clock and max fsb and let the bios calculate the multiplier internally if necessary.
ie. 2764mhz / 251mhz = 11.011952191235059760956175298805 =>11.012.
Absurd?

 

[redhatlinux]

Run the higher FSB with relaxed memory timings. Forget the benchmarks, there are virtually no real world applications which stress memory that much. Your gain in CPU performance will be a benefit to every application.

 

[pallmall]

Originally posted by: redhatlinux
Run the higher FSB with relaxed memory timings. Forget the benchmarks, there are virtually no real world applications which stress memory that much. Your gain in CPU performance will be a benefit to every application.

That's why most of the memory suppliers try to sell the expensive memory with higher clock rather than inexpensive memory with fast mem timings.
But some benchmarks indicate that the effect of fast mem timings in some condition exceeds that of higher clock.
Rather than just rushing to get the higher FSB or faster mem timings, somebody may need some background knowledge before or after the accident.

To check the trade off of the performance factors, I'd like to list them by my priority:
1. CPU clock : multiplier * FSB : higer is better of course
affects the speed of actual calculation, data transfer between cpu and L1,L2 cache
2. FSB : higher is better
affects the communication speed between cpu and north bridge including memory, agp and indirectly with south bridge and its attached devices
3. Memory clock : higher is better
4. Memory timings : lower is better
5. other individual devices

There is no doubt that 1 and 2 must be higher.
But for 3 and 4, there are actually small or no difference as far as the memory clock and timings are both fully or equally optimized when compared (IT IS HARD).
The confusions mostly come from focusing on only 1 or 2 factors, not considering all major factors(1 THRU 4).

 

[irabufan]

thanks for the input
i wanted to run my system at a higher fsb if i could because a 14x multiplier seems kinda high
and i haven't seen anyone around here using such a high multiplier

 

[pallmall]

If you think your cpu is maxed out around 2332mhz, you can lower the multi to 10.0 or 10.5.
Somebody said your ram can do over 230 with relaxed timings such as 7,3,3,3.0, of cource with higher vmem.
You can even try higher fsb with async mode like 5:4 ratio if the nb and cpu supports.
Or you can try the same fsb with lower memclock with faster mem timings if you want.
It is very hard to optimize the mem timings for the mem clock because of their relatively few(coarse) options.
Use benchmark programs such as memtest86, aida, sandra, superpi or 3dmark2001 for fine comparisons or optimizations.
Abit forum or search in this site will help you for more references.