What are best RAM timings to lower?

[Metalloid]

I heard that lowering the Cas latency doesn't help much. What settings should I lower?

 

[Lizardman]

see thugsrook cas shootout.

 

[Metalloid]

Ok thanks that does help, but I have a question. What are the 4 cas timings?

I know that latency is first, but what is the order for the next 3?

 

[nania]

Metalloid15
Memory timings can be expressed 2 2 2 5T or T6-2-2-2 and both mean the same thing. It would be nice if the number was presented in a consistent format because as we get to faster memory, the potential for confusion will increase. The memory timings (CAS latency) have a significant impact on perfomance at lower clock speeds but as the CPU gets into the multi-GHz zone, the bandwidth of the memory becomes more important. That is why you have the notion that relaxing the latency doesn't hurt your performance because in a 3GHz P4, the internal parasitics of the CPU limit the advantage of the faster latency but the bandidth required in this example is very important. There is a point where the CPU is limited by the bandwidth and latency of the memory and that point is best identified by the multiplier. Identifying the multiplier was the reason for this thread.

 

[nania]

edit:
...2 2 2 5T or T5-2-2-2
sorry

 

[chizow]

Originally posted by: nania
Metalloid15
Memory timings can be expressed 2 2 2 5T or T6-2-2-2 and both mean the same thing. It would be nice if the number was presented in a consistent format because as we get to faster memory, the potential for confusion will increase. The memory timings (CAS latency) have a significant impact on perfomance at lower clock speeds but as the CPU gets into the multi-GHz zone, the bandwidth of the memory becomes more important. That is why you have the notion that relaxing the latency doesn't hurt your performance because in a 3GHz P4, the internal parasitics of the CPU limit the advantage of the faster latency but the bandidth required in this example is very important. There is a point where the CPU is limited by the bandwidth and latency of the memory and that point is best identified by the multiplier. Identifying the multiplier was the reason for this thread.

That's the case for a P4 system, but Athlon XP rigs are almost the exact opposite. XP's can't fully utilize available bandwidth like a P4, so sacrificing timings for the sake of higher bandwidth can lead to latency penalties. Any increase in performance gained from extra bandwidth can be negated and offset by higher latency. I think there are exceptions to this rule as well, as someone pointed out that servers and DC rigs benefit from higher bandwidth even if it means sacrificing some in the form of tighter RAM timings.

Chiz

 

[nania]

Chizow

That's the case for a P4 system, but Athlon XP rigs are almost the exact opposite.

Are you saying that the Athlon XP is immune to increasing internal parasitics, immune to bandwidth limitations or both? If it works the "exact opposite", the Athlon could scale infinitely with the 133FSB. If your statement is correct, there is no optimum multiplier for the Athlon XP. Please clarify your statement.

 

[chizow]

It works the opposite with an XP, as it doesn't benefit from faster FSB/increased bandwidth as much as it benefits from faster RAM timings. Since XP's still have yet to reach the clockspeeds necessary to benefit from much more than a 333MHz FSB, running a 400MHz FSB to increase bandwidth but sacrificing RAM timings leads to a degradation in performance due to increased latency. Also, XP's have been unlocked or unlockable for some time, so yes, there is no optimum multiplier on any given chip. The optimum multiplier is whatever you set to achieve the highest sustainable clock speed to match the fastest FSB you can obtain w/out sacrificing RAM timings.

That doesn't mean that the increased bandwidth isn't beneficial in some instances, but only if RAM timings can be kept low and the total clockspeed can benefit from the increased bandwidth. Also, its very difficult in comparing between the P4 and Athlon b/c there are considerable differences in their architecture, not to mention the options an Athlon user has in terms of controlling all the variables mentioned above. It seems you are referring to set multipliers and FSB intervals; with an AMD system on a KT400 or nForce2 platform, these limitations don't exist.

Chiz

 

[nania]

Chizow

XP's have been unlocked or unlockable for some time, so yes, there is no optimum multiplier on any given chip. The optimum multiplier is whatever you set to achieve the highest sustainable clock speed to match the fastest FSB you can obtain w/out sacrificing RAM timings.

Let's assume we are using DDR266 (ie: PC2100 DRAM) a 2 2 2 6T. At what clock speed will the Athlon XP begin to be limited by its memory accesses. In other words, where will an increase in CPU speed begin to bring a disproportially lower increase in performance.

 

[nania]

...or conversely, at what point do we stop seeing increases in performance with just raising the memory bus (ie: lowering the multiplier)

 

[Metalloid]

Ok, would someone like to please explain to me why lowering my RAM timings has actually given me a lower score in 3DMark03? I have so far run in default (2.5,3,3,6); 2,3,3,6; 2.5,2,3,6; and 2.5,3,3,5; and the default has had a score that is about 100 points higher every time. I thought lowering RAM timings was supposed to yield better performance.

 

[John]

Originally posted by: Metalloid15
Ok, would someone like to please explain to me why lowering my RAM timings has actually given me a lower score in 3DMark03? I have so far run in default (2.5,3,3,6); 2,3,3,6; 2.5,2,3,6; and 2.5,3,3,5; and the default has had a score that is about 100 points higher every time. I thought lowering RAM timings was supposed to yield better performance.

Run something besides 3dmark to bench your memory performance.

 

[Metalloid]

Originally posted by: John

Originally posted by: Metalloid15
Ok, would someone like to please explain to me why lowering my RAM timings has actually given me a lower score in 3DMark03? I have so far run in default (2.5,3,3,6); 2,3,3,6; 2.5,2,3,6; and 2.5,3,3,5; and the default has had a score that is about 100 points higher every time. I thought lowering RAM timings was supposed to yield better performance.

Run something besides 3dmark to bench your memory performance.

Ok I will try other things, but I was hoping this would give me an idea of how much faster my 3D apps will be.

 

[iLoveDivX]

bumpity for you