PC2700 w/Low Cas vs. PC3200 w/High Cas?

[groovin]

i know there could be lots of variables in this, but lets assume no overclocking, PC2700 with cas latency of 2-2.5 and PC3200 with cas of 2.5-3. i already have some pc2700 RAM like this, but i was wondering if id see any improvement moving to pc3200 of this kind.

thanks

 

[Hajime]

What platform?

P4, AXP, A64...P4 mobile? P3? Original Athlon? Or even Via's C3?

 

[groovin]

Athlon XP 2800 on nForce II

 

[formulav8]

Run the memory in sync with the fsb whatever the fsb is. Then maybe worry alittle more about timings.



Jason

 

[Matthias99]

Originally posted by: groovin
Athlon XP 2800 on nForce II

Platform is irrelevant for comparing raw memory speeds.

PC3200 CL3 = 200Mhz (400Mhz effective) = 5ns cycle time. CL3 = 3 cycles / access. 3 * 5ns = 15ns / access.

PC2700 CL2.5 = 2.5 * 6ns = 15ns / access.

PC3200 CL2.5 = 2.5 * 5ns = 12.5ns / access.

PC2700 CL2 = 166Mhz (333Mhz effective) = 6ns cycle time. CL2 = 2 cycles / access. 2 * 6ns = 12ns / access.

In terms of latency, PC2700 CL2.5 is exactly equivalent to PC3200 CL3, and slightly slower than PC3200 CL2.5. PC2700 CL2 is faster than PC3200 CL2.5, but just barely.

In terms of bandwidth, the PC3200 will blow the PC2700 away (it's about 20% faster).

On an AXP-based platform, you generally take a pretty big performance hit running your memory async. So if you have a 166Mhz FSB and you're not overclocking it, stick with the PC2700. Athlon processors generally are not starved for bandwidth like P4's are. If you have a 200Mhz FSB, you should run at PC3200 speeds to avoid running async, even if you have to loosen the timings a bit. Frankly, I doubt you'd see much difference outside of a few synthetic benchmarks -- the difference between the tighest and loosest timings is generally well under 5% in practice.

 

[Hajime]

Matthias99, only one comment there.

What about DDR2 800 to DDR2 1066 on the P4? Eh? EH?

Looking at the benchmarks, tighter timings may very well lead to a slightly faster machine overall - even if it's only 5%, what if 1066 only nets a 3% increase?

Just a bit of food for thought.

And while you take a big hit for running your ram asynch, AXP's don't blow chunks by suddenly going from a 200fsb to a 166fsb. Also, if memory serves, a surprising amount of AXP's are multiplier unlocked and AXPs are -hardly- bandwidth starved - hence, it is very possible that decreasing the fsb/increasing the mult/tightening the timings may very well give better results.

Of course, YMMV. In my case, two times doing that led to better AXP benchmarks. But one time it didn't.