Should I buy ddr2 1200 memory since I'm going to overclock my system anyway.

[mcduke]

Well, I'm taking the plunge and going to upgrade present system (since it's about ready to be used as a paper weight). I'm also looking to do some overclocking. Something I've had limited success with in the past. I'm looking to get the following:

Intel Core 2 Duo E6750 Conroe 2.66GHz 4M shared L2 Cache LGA 775 65W Dual-Core

EVGA 122-CK-NF67-T1 LGA 775 NVIDIA nForce 680i LT SLI ATX Intel Motherboard

And for memory I was thinking about getting,
Crucial Ballistix 2GB (2 x 1GB) 240-Pin DDR2 SDRAM DDR2 1066

I figured if I was going to overclock the system that I would just get memory that's already rated faster then the ddr800.

My son in law says that I should just a get a proven ddr800 overclocker for memory. That getting the ddr2 1066 doesn't work well with the cpu with an overclocked system. This doesn't make sense to me. Can anyone explain what I should be looking at for memory when overclocking and why I shouldn't be looking at the ddr2 1066 or even the ddr2 1200?

 

[JamesXP]

You don't even need high speed ram with 680i, You can run the memory at a different speed than the FSB also, I'd get a 780i board if you need SLi, If you don't, Get a good P35/X38 board.

 

[Bageland2000]

to the forums, stay a while. For starters, I suggest reading up on the basics of overclocking in the stickies, but if you don't end up doing that here is a quick reference. Your CPU (e6750 in your case) and RAM usually are run at the same speed in terms of mhz. your 6750 runs at 333mhz (i know it's 2.66ghz but its front side bus is actually at 333x8 to get 2.66GHZ)) so your processor is technically at 333mhz. ddr2-800 runs at 400mhz (which will automatically run at 333mhz when your processor is in) if you overclock past 400mhz on your CPU you will overclock you RAM too unless you get ddr2-1066 which runs at 533mhz or ddr2-1200 which runs at 600mhz. I'm not sure where your brother-in-law is coming from when he says that 1066 doesn't run well with overclocked processors, maybe an isolated experience that he had, normally ddr2-1066 should run better that 800. Anyway thats a quick (possibly confusing explanation) and like I said before, read up on overclocking. Hope that helped

 

[Albuquerque]

Just to reiterate and perhaps restate what Bageland2000 mentioned...

Intel processors have a front-side bus that is a bit misleading. Intel says "1333mhz front side bus", but the reality is actually 333mhz clock rate with four data transmissions per clock. Hence, you get an effective data rate of 333 x 4 = 1333mhz.

Double Data Rate (DDR) memory works in a similar fashion -- DDR2/800 works at 400mhz clock rate with two data transmissions per clock. Hence, you get an effective data rate of 400 x 2 = 800mhz. Rinse and repeat for all other speeds, examples: DDR2/1066 (533 x 2) or DDR21200(600 x 2)

Next-to-last thing: Essentially every consumer motherboard on the market today has a dual-channel memory interface. This essentially boils down to the motherboard being able to read from two banks of memory simultaneously, thus doubling memory bandwidth in nearly all circumstances.

Very last thing: no current intel processors have an integrated memory controller like the AMD's do. All memory interface is done from the chipset northbridge, so it goes like this: CPU -> Northbridge -> RAM. When you select a certain memory speed, that will be the speed of the link between the northbridge and ram. When you select a certain FSB speed (ie, overclocking the CPU), that will be the speed of the link between the CPU and the northbridge.

Ok, are you ready for the conclusion?

FINAL EXAM ANSWERS
The link between your CPU and northbridge is determined by the FSB.
The link between your memory and northbridge is double the speed of the memory if you're using two sticks of ram (dual channel)
There is also a "ratio" component that determines how fast your memory goes versus the FSB. The only one you need is 1:1, and if you have questions, see the footnotes to this post.

So... Really really simple math time:
FSB = Memory speed * number of memory channels used (1 or 2 are the only options)

Examples of dual-channel motherboards and the necessary memory speed:
FSB: 1333Mhz
Memory: 667Mhz

FSB: 1600Mhz
Memory: 800Mhz

FSB: 2000Mhz (this is pretty close to the limit for most boards off-the-shelf)
Memory: 1000Mhz

So, in order to get the true benefit of your DDR2-1200 memory, you'd need to run an FSB of 2400Mhz. Unless you plan on some SERIOUS hardware (both in terms of cooling, modifications to the board, and power) then you'll never get your use out of them.

Footnote:
Many motherboards allow for memory speeds that are OVER the FSB speed, such as 5:6, 3:4, even 1:2. These technically allow the memory to operate at a faster clockrate than the front side bus speed. So then why doesn't everyone buy DDR2/2000 and have uber benchmark scores? Because if you're using more than one memory stick, it's effectively worthless

Here's why: Remember the FSB? This is the bus linking the CPU to the northbridge. All memory access MUST go through the northbridge on Intel processors (currently; this will change with Nehalem in '08).

So, for any memory access to happen, it MUST come from the CPU, across the front side bus, to the northbridge, and then to memory. Herein lies the problem: if you're running dual channel, then the front side bus becomes the bottleneck.

Think about it:
Front side bus at 1600mhz (it's a 400mhz clock rate underneath)
Memory speed at 800mhz (it's also a 400mhz clock rate underneath)
Two memory channels running in dual-channel mode.

In the above scenario, you capable of ENTIRELY saturating the front side bus with memory traffic. If you raise the memory speed to 1200mhz, it gains you nothing, because 1200 x 2 = 2400 which is still less than the 1600mhz front side bus.

You could raise the memory speed to a gazillion gigahertz, but so long as your front side bus is still at 1600mhz, your system speed really isn't going to budge. There are corner cases where you can see incredibly tiny gains (two percent or less) in tasks that do a LOT of very organized and linear memory address reads, such as un-raring absolutely massive files from within ram. This comes from the northbridge having a small amount of prefetch logic for these sorts of cases, but it's also offset by the lack of buffer and latencies in the asynchronous signal timings.

 

[RockerGuy]

This comes from the northbridge having a small amount of prefetch logic for these sorts of cases, but it's also offset by the lack of buffer and latencies in the asynchronous signal timings.

Great write up.

I love ur system in ur sig. too.
Its pretty tricky overclocking old system, no?? I am a noob, so what do I know.

 

[mcduke]

Wow, I think my brain hurts from reading all that. My present system is an amd Opteron dual core 2.0 (overclocked to 2.1) on a 939 motherboard, which has the old ddr 400 memory (and because of the way amd has their cpu setup, by ovcerclocking the cpu the memory is also overclocked). Unfortunately, that is all I could get out of my present system (apparently a bum cpu).
The motherboard I'm getting runs a 1333 fsb. Now from what my son in law told me about intel 680i motherboard, the cpu can be overclocked separate from the memory. He's running his 2.0 cpu at 2.8 (with just standard cooling and no problems).
But, I wanted to overclock the cpu and memory and that is why I figured I'd get the ddr2 1066 memory. From everything I've read, and am trying to understand, It sounds like the highest I should consider is the ddr2 1066? Then of course what about the timing on the memory. I guess I was just hoping for a simple answer of - get this and you'll be all set.

 

[Albuquerque]

Regardless of what a large quantity of fanbois may suggest, run your memory at 1:1 for maximum performance, stability and power efficiency. So if you're targetting for something near 500FSB, then DDR2/1066 would be a good purchase. If you're more likely to max out around 400-450FSB, then some high quality DDR2/800 may be a better option.

Timings are indeed important, so do your best to find something with decent timings if you can. But I wouldn't worry about anything faster than DDR2/1066.

 

[jdmnash]

... computer enthusiast-ing for dummies, part 1. great write-up!

(will proceed to read 6 more times or whenever I "get it")

((who am i kidding!?!? will read 6 more times))

 

[jmsanders2]

Regardless of what a large quantity of fanbois may suggest, run your memory at 1:1 for maximum performance, stability and power efficiency. So if you're targetting for something near 500FSB, then DDR2/1066 would be a good purchase. If you're more likely to max out around 400-450FSB, then some high quality DDR2/800 may be a better option.

Timings are indeed important, so do your best to find something with decent timings if you can. But I wouldn't worry about anything faster than DDR2/1066.

Right on as usual, great write up Albuquerque. My opinion doesn't hold as much weight as Albuquerque's, but I agree and it depends on what you are trying to OC to...the e6750 has a x8 multi and so if you are trying to go above 3.6ghz I would say you're in the ddr2 1066 region to run 1:1. Really it's a matter of cost vs. what you are trying to oc to. You could probably hit above 450 fsb with ddr2 800 running 1:1, but you would probably have to lower the timings.

 

[Hazaro]

The only reason to buy 1200 would be to tighten them at 800 or 1066 speeds. And the only mobo that I know of currently that can hit 600 FSB is a modded IP35 Pro.

Basically just go with 800 or 1066 DDR2

 

[keninishna]

I say faster is always better . I think memory cycles do not exactly equal CPU cycles. Other things like memory timings play a role. I believe the memory requires several cycles to fetch data and all those cycles your memory is going through your CPU is waiting on that data because I think the CPU only requires 3 cycles to request to read from memory? So your CPU spends 3 cycles requesting the read for instance and it takes your memory 7 cycles to actually put the data into the CPU cache. So your CPU is wasting 10 cycles every memory read. If you clock the memory to twice the speed of the CPU then those 7 cycles the memory spends is actually equivalent to 3.5 cpu cycles which would reduce 3 CPU cycles for every read. Someone feel free to correct me if I'm wrong. For some reason windows seems much snappier for me with a 4:5 ratio versus 1:1

Here is 1:1
http://farm3.static.flickr.com/2278/2164867284_d22610848f_b.jpg
and 4:5
http://farm3.static.flickr.com/2214/2164867290_e3ded309b7_o.jpg

 

[Albuquerque]

I say faster is always better . I think memory cycles do not exactly equal CPU cycles. Other things like memory timings play a role. I believe the memory requires several cycles to fetch data and all those cycles your memory is going through your CPU is waiting on that data because I think the CPU only requires 3 cycles to request to read from memory? So your CPU spends 3 cycles requesting the read for instance and it takes your memory 7 cycles to actually put the data into the CPU cache. So your CPU is wasting 4 cycles every memory read. If you clock the memory to twice the speed of the CPU then those 7 cycles the memory spends is actually equivalent to 3.5 cpu cycles. Someone feel free to correct me if I'm wrong. For some reason windows seems much snappier for me with a 4:5 ratio versus 1:1

Here is 1:1
http://farm3.static.flickr.com/2278/2164867284_d22610848f_b.jpg
and 4:5
http://farm3.static.flickr.com/2214/2164867290_e3ded309b7_o.jpg

Want to guess why? Here's a hint: SINGLE CHANNEL. Look at the screenshots you posted again. If you're running single channel, then you'd do even better to run it at 1:2 ratio versus the 4:5 you're running now. The reason? Because your memory bus has far less bandwidth in single channel mode @ 1000mhz than your FSB does at 1600mhz. If you were in dual channel mode, then it would be a different story.

Now, back to your timings question -- while timings do play a part, it goes both ways. If the clocks are asynchronous (the busses are not at the same signalling rates) then you're also going to have cases where requests in both directions are waiting for the signals to line up. Further, in dual channel mode, you're half as likely to run into full-latency waits because the dual memory channels can be on different timing sequences.

 

[keninishna]

hah, **** I thought something was off. It seems with dual channel bandwidth memory cycles effectively are less than or equal to those of the CPU cycles.

 

[Albuquerque]

hah, **** I thought something was off. It seems with dual channel bandwidth memory cycles effectively are less than or equal to those of the CPU cycles.

Not exactly sure what you meant by the second part, but essentially in dual channel it's pointless to exceed 1:1, and it's performance-detrimental to drop below 1:1.

 

[keninishna]

That was weird I just reseated my memory sticks and they are back in dual channel mode.

Anyway what I'm thinking is the latency for the memory clocks are determined by the memory bus. So even if the memory is running at 2400 mhz it would get divided by the latency. So CAS 5 which would slow down the data transfer to 480mhz on the fsb. But apparently the latency clocks are based on the fsb and not the memory bus. I'm too tired now to think, I'll just research it tomorrow but you are right as the benchmarks tell the truth.

 

[Albuquerque]

No, you had the right idea the first time -- timings are in relation to the NB -> MEM bus, not the CPU -> NB bus. But again, since the northbridge can access and run dual channels seperately but simultaneously, a specific request has a 50% greater chance of being on a channel that's already being fetched in dual-channel mode.

While it's possible to "artificially" reduce effective timing by way of mem clock above FSB, you still face the latency problems inherent with asynchronous busses trying to communicate. Even at 1:2 ratio where the memory is double the speed of the FSB so you get maximum potential timing effect, you'll still have the problem with data coming back from memory waiting for the FSB. The northbridge has a tiny amount of buffer for some minor prefetching, but it's not going to last long at that a double-rate.

 

[keninishna]

I think I understand now even though the memory is faster clocked it still has to line up with the fsb clock. So in the end its still gonna go the same speed or slower than the fsb.
If thats the case I wonder why motherboard makers even support memory speeds faster than the fsb? I guess if you stick ddr1200 ram on a 333mhz bus it might not even boot cause it would be running at like half the rated speed.

 

[GruffiGummi]

so with an amd 6400 black edition at 3.2 (200 x 16) my ddr2-800 ram is only running at 200? or is it 400

 

[Albuquerque]

Memory has been available that's faster than FSB ever since the original 845 chipset. The reason why is because of single-channel users. In the beginning, single channel was the only option. Thus, you had far more FSB bandwidth than memory bandwidth; Intel wanted to make up for it somehow.

So, if you've got an FSB-1333 system, then you can run a single DDR2/1333 memory stick and get essentially the same benefit as a dual channel pair of DDR2/667 sticks. Is that a good tradeoff? Each user would have to decide for themselves. But for big manufacturers, running a single cheap DDR2/1066 chip might be cheaper than running two DDR2/533 sticks of half-size -- while still getting the same perf. Make sense?

 

[keninishna]

true I didn't think of that. Too bad two 512 sticks are usually cheaper than a 1 gig stick at a faster speed.