Memory Latency vs. Frequency

[SweetAndLow]

So I have found some information about this subject but wanted to ask on here since file server workflows might be different than the benchmarks everyone else is using. I'm working on putting a build together and I need to decided between 16G crucial 1333 CL9 unbuffered ECC memory and 16G 1600 CL11 unbuffered ECC memory. As of right now it seems as though they are dead even performance wise. Does anyone have any other insight into memory selection?

Resources:
http://www.anandtech.com/show/7364/memory-scaling-on-haswell/6
http://www.thetechrepository.com/showthread.php?t=160

 

[warri]

I doubt that you are going to notice any difference. Just buy the ones for which you know that they work with your mainboard ;)

 

[jyavenard]

The article you posted show that 1600MHz no matter the latency having a slight edge under all tests over 1300MHz DD3.

Having said that, as mentioned, I doubt you'll notice a difference.
FWIW, I've just built a system for a friend, the 8GB 1600MHz memory was cheaper than the 1300MHz equivalent and the 8GB dimm was only 25% more expensive than the 4GB one.
So went with 8GB/1600MHz purely for the GB/$ perspective...

 

[cyberjock]

Latency really doesn't matter that much. Obviously if its $5 for a faster stick that's probably a better deal, but be smart with it. Money should be spent to increase your amount of system RAM over timings.

If you do some cool math you'll see something:

1333Mhz memory at CL9:

1/(1333000000)= 7.5x10^10 (this gives you the length of time per clock cycle. In this case, roughly .75nanoseconds per clock cycle

7.5x10^10 x 9 = 6.75ns

So your RAM takes 6.75ns for any read or write.


1600Mhz memory at CL11:

1/(1600000000)= 6.25x10^10 (Roughly .625nanoseconds per clock cycle)

6.25x10^10 x 11 = 6.875ns

So your RAM takes 6.875ns. for any read or write.


So, which is faster? 6.75ns(1333 memory) or 6.875ns(1600 memory)?


Just by checking these numbers out, its likely that your 1333 memory could run at 1600 CL11 if you manually set those values since its faster by clock cycle but lower by actual latency. I say "likely" because the clock cycle does have an effect on the ability to reliably write, read, and store the data.

This is how many companies cheat. They give you very high Mhz ratings, but the latency is so high that "slower" chips are technically better. There's other factors that muddy these waters even more, but I think I've made my point. Don't assume that ddr3-1600 must be faster than 1333. ;)

 

[Dusan]

So, which is faster? 6.75ns(1333 memory) or 6.875ns(1600 memory)?

This is not so simple. The times you calculated only apply for the first transferred word. However, current CPUs use burst transfers and the time to transfer the subsequent words depends also on the data transfer rate (where 1600 is faster than 1333). Check the table (and explanation) here (especially the first/fourth/eight word columns for DDR3): https://en.wikipedia.org/wiki/CAS_latency

EDIT:

1333Mhz memory at CL9:

1/(1333000000)= 7.5x10^10 (this gives you the length of time per clock cycle. In this case, roughly .75nanoseconds per clock cycle

7.5x10^10 x 9 = 6.75ns

So your RAM takes 6.75ns for any read or write.

You also made a mistake here. DDR-1333 does not run at 1333MHz. The SDRAM chips are clocked at 167MHz and the I/O bus runs at 667MHz, so the correct result is 13.5ns.

 

[cyberjock]

Doh. Those numbers seemed a little low when I did them. I just figured I was remembering something wrong. Glad someone caught the mistake.

I agree with the wikipedia explanation, but its actually more complex than even that webpage makes it out to be. The refresh rate and a bunch of those other values besides just the CAS latency do have an effect. I've been unable to validate the actual quantity of the affect since these concepts become very complex and generally aren't discussed with the public.

The point I was trying to make is that frequency isn't the complete truth for how a memory stick will perform. Hopefully I got that point across. ;)