Solution
The current Intel nehalem and sandy bridge cpu's have an excellent integrated ram controller. It is able to keep the cpu fed with data from any speed ram.
The difference in real application performance or FPS between the fastest and slowest ram is on the order of 1-3%.

Synthetic benchmark differences will be impressive, but are largely irrelevant in the real world.

Fancy heat spreaders are mostly marketing too.

Only if you are seeking record level overclocks should you consider faster ram or better latencies.
Read this Anandtech article on memory scaling:
http://www.anandtech.com/show/4503/sandy-bridge-memory-scaling-choosing-the-best-ddr3/1
---------------bottom line------------

DDR3 1600 is the sweet spot.

I see no negative...

slicedtoad

Distinguished
Mar 29, 2011
1,034
0
19,360
assuming it's for gaming with an ivy/sandy cpu.
8GB 1600.

Speed past 1600 offers less than a 1% increase in performance.
16GB is useless unless it's a workstation. In that case the 16GB 1866 would be best.

If had to go with one of the above for gaming I'd choose the 16GB though. In a 4x4GB config, cause it looks nice to have all the slots filled up. And ram speed can't be made use of past 1600mhz.

All this would change if we moved away from ivy/sandy. I believe lano see considerable advances from ram speed since the on-die gpu uses that ram.
 

Nope, bad idea. Speeds over 1600 offer ZERO performance benefits. You're literally throwing $ away.
 
They are all right!

DDR3-1600 is the "sweet spot" for Sandy Bridge CPU's, I assume the same for Ivy Bridge. Any bump in spped will not be noticed, only cost you $$.

If you (can) overclock it, then ok, all the better.

No real difference between CAS/CL 8 or 9, either.
 
The current Intel nehalem and sandy bridge cpu's have an excellent integrated ram controller. It is able to keep the cpu fed with data from any speed ram.
The difference in real application performance or FPS between the fastest and slowest ram is on the order of 1-3%.

Synthetic benchmark differences will be impressive, but are largely irrelevant in the real world.

Fancy heat spreaders are mostly marketing too.

Only if you are seeking record level overclocks should you consider faster ram or better latencies.
Read this Anandtech article on memory scaling:
http://www.anandtech.com/show/4503/sandy-bridge-memory-scaling-choosing-the-best-ddr3/1
---------------bottom line------------

DDR3 1600 is the sweet spot.

I see no negative with 16gb. No game will use more than 2-3gb, but lots of ram will be used by windows to keep code in ram, ready for instant launch.

Ram is sold in kits for a reason.
Ram from the same vendor and part number can be made up of differing manufacturing components over time.
Some motherboards can be very sensitive to this.
Although, I think the problem has lessened with the newer Intel chipsets. Still,
it is safer to get what you need in one kit.

It often turns out that a single kit will be cheaper too.
 
Solution
8gb sticks have come down in price, but are still a bit more expensive per gb than 4gb sticks.

The advantage of 16gb using 2 x 8gb is that the voltage to run two sticks is easier to manage than 4. When oc'ing, this may be an issue, depending on the motherboard.

2 x 8gb also gives you an option to go to 32gb.

Ram is relatively cheap, so it does not pay to anguish too much about it.

You want documented ram compatibility. If you should ever have a problem, you want supported ram.
Otherwise, you risk a finger pointing battle between the ram and motherboard support sites, claiming "not my problem".
One place to check is your motherboards web site.
Look for the ram QVL list. It lists all of the ram kits that have been tested with that particular motherboard.
Sometimes the QVL list is not updated after the motherboard is released.
For more current info, go to a ram vendor's web site and access their ram selection configurator.
Enter your motherboard, and you will get a list of compatible ram kits.
While today's motherboards are more tolerant of different ram, it makes sense to buy ram that is known to work and is supported.