Memory for X79/3960x

[cm123]

I've heard that some people that went 32GB on X79 (8gbx4) has actually made their systems feel a bit sluggish vs 16Gb, so bit concerned about jumping up to 32GB.

I'm thinking about also 1866MHz - G.Skill 8-9-9-24 4GBx4 - like low latency

http://www.newegg.com/Product/Produc...82E16820231539

or

Considering is go 32GB Crucial 8GBx4 1600MHz 8-8-8-24

http://www.crucial.com/store/mpartsp...101CC8A5CA7304

also thinking about Crucial 4GBx4 1600MHz 8-8-8-24

http://www.crucial.com/store/mpartsp...B5CAE5A5CA7304

While it be nice to have (way more than I need to mostly game and some graphics work) 32GB, I'm concerned about the laggy comments, also any problems with i7-3960x and Asus SABERTOOTH X79?

What does everyone think?

 

[aaksheytalwar]

Since you have money to burn,

I recommend getting DDR3 2400+ low latency RAM. 16 or 32 as you please.

Though I would suggest a 3930k or 3770k with 16gb Samsung 30nm ram instead if you upgrade only once in 1-2 years or less and don't want to waste money for zero performance gain

 

[cm123]

I'm not sure want to overclock much, 1866MHz is about my limit - doesn't the latency get pretty high on the 2133 and 2400+ taking away much of the gains?

One of my other systems is i7-2700k, I think it feels more responsive than this system 3960x) in FPS games myself - I have the 3960x and really just looking to get the most bang out it.

Since you have money to burn,

I recommend getting DDR3 2400+ low latency RAM. 16 or 32 as you please.

Though I would suggest a 3930k or 3770k with 16gb Samsung 30nm ram instead if you upgrade only once in 1-2 years or less and don't want to waste money for zero performance gain

 

[cmdrdredd]

Latency decreases with frequency as well.

 

[jwilliams4200]

Note that latency is usually specified in IO clock cycles, NOT nanoseconds (ns).

For 1866 memory (note, NOT 1866MHz, many people make the mistake you did), that is 1866 MT/s, but it is double-data-rate (DDR), so the IO clock is 1866/2 = 933MHz. That means one cycle is 1/933MHz = 1.07 ns.

If it is CAS9 latency, then the actual latency in nanoseconds would be 9x1.07 = 9.6 ns.

For comparison 1600MT/s memory with CAS8 latency would be 8x1.25ns = 10 ns.

 

[cm123]

Thanks - I'll change direction all together and get DDR-1600MHz 7-7-7-24 (4GBx4)...

Note that latency is usually specified in IO clock cycles, NOT nanoseconds (ns).

For 1866 memory (note, NOT 1866MHz, many people make the mistake you did), that is 1866 MT/s, but it is double-data-rate (DDR), so the IO clock is 1866/2 = 933MHz. That means one cycle is 1/933MHz = 1.07 ns.

If it is CAS9 latency, then the actual latency in nanoseconds would be 9x1.07 = 9.6 ns.

For comparison 1600MT/s memory with CAS8 latency would be 8x1.25ns = 10 ns.

 

[mrpiggy]

I have a Asus P9X79 Pro with a i7-3960x and 32GB Crucial 1333RAM (8x4GB). If you want the x79 and 3960x to be a gaming system, I suggest turning off hyperthreading. It seems more responsive in games. If you actually do do lots of very intensive things at one time, then the hyperthreading makes sense and having the OS see 12 cores helps with those intensive HT-aware applications.

For a gaming machine, you don't need 6 extra virtual cores when you have 6 physical cores and a shitload of cache like these 3960x's have. Not having to deal with hyperthreading makes it run faster and cooler (whether you OC or not). Someone (forgot where/who on internet) actually tested the 3960x processor power/throughput with HT on versus off and the GFlops throughput was higher with HT off. Sure it's not as efficient 12 cores if you are running lots of software that can thread through multiple cores, but most of us NEVER run anything other than benchmarks that can even load down 6 cores, and having less OS background tracking/decision making without HT on seems to help in game performance with this CPU.

As for memory speeds, I have another setup exactly like this at work with the only difference being faster memory. It shows no difference in anything other than certain specific benchmarks and very, very little at that. Overclock your system 10% and you get 10% more processor performance in everything. Use even the fastest memory and you are lucky to get 2% difference even in benchmarks and no felt difference at the keyboard/mouse. Not saying not to get faster memory if the price is the same, but I would not sweat 1333 versus whatever in an x79 board. Don't forget that the x79 is quad-channel memory setup versus dual channel of the normal desktop chipsets (z77, etc). Faster memory does not seem to affect the x79 quad channel as much as it affects the performance of dual channel configurations. I addition, the 3960x has a ton of internal cache that masks RAM performance.

 

[cm123]

DDR3-2133MHz 11-11-11-27 - how many ns then does this translate to?

Could you run DDR3-2133MHz at DDR3-1600MHz with lower cas like 7-7-7-24?

Note that latency is usually specified in IO clock cycles, NOT nanoseconds (ns).

For 1866 memory (note, NOT 1866MHz, many people make the mistake you did), that is 1866 MT/s, but it is double-data-rate (DDR), so the IO clock is 1866/2 = 933MHz. That means one cycle is 1/933MHz = 1.07 ns.

If it is CAS9 latency, then the actual latency in nanoseconds would be 9x1.07 = 9.6 ns.

For comparison 1600MT/s memory with CAS8 latency would be 8x1.25ns = 10 ns.

 

[jwilliams4200]

DDR3-2133MHz 11-11-11-27 - how many ns then does this translate to?

Could you run DDR3-2133MHz at DDR3-1600MHz with lower cas like 7-7-7-24?

IO Clock = 2133/2 = 1066.5MHz
IO cycle time = 1/1066.5MHz = 0.9376 ns
latency = 11x0.9376 = 10.3 ns

As for running 2133 CAS11 at 1600 CAS7, all anyone can say is "maybe". The memory is certified to run at the given specifications. Anything else you have to try and see.