Hi all,

I have reposted this from the Memory forum as it seemed to not be getting much attention. Sorry in advance.

I am new to Tom's Hardware and a bit of a n00b to hardware specs and how they all fit together and provide "true collective" vs. "stock individual component" performance. Sorry for my made up terms if they are just plain daft and better bona fide terminology exists out there for these conecpts.

I have been reading quite a bit and so have tried to "self help", but there is one thing I cannot seem to find a definitive answer to. Let's say you get the Intel E8500 which has a speed of 3.16GHZ and an FSB of 1333MHZ. Being Intel, this is quad pumped and so we have a system clock of 333MHZ making the internal multiplier 9.5 to give you the stock speed of the processor. All this I understand.

So we now want to get memory that can make the most use of our dandy new CPU. We are going with DDR, obviously. DDR is dual rate, so 2 x system clock which, as we calculated earlier, is set to 333MHZ = 667MHZ. Now we can see that there is DDR-3 1333MHZ available nowadays that matches the 1333MHZ of the CPU's FSB. Ahhhh, but hang on. Even though the memory can run at this speed it is still only DDR meaning 2 x clock speed. Correct? This IS my question. At the moment I understand the answer to this question to be "yes", which means the memory will run at 667MHZ. So, latency issues aside, why not just go for DDR2-667 or DDR2-800? Of course, I realise the 800 will run slower, but I read that even though it won't be 20% faster it will be marginally better than the step below. Is this correct? If so, then I may as well do it as the price gap is negligible in my estimation.

But back to the original question. Apparently, the ideal is to do a 1:1 so you are getting the most out of the "4x" of your processor's and "2x" of your memory's capabilities. So, if you have 1066MHZ processor FSB, get 533MHZ RAM ("4x" / "2x" ). Faster memory will provide a "diminishing returns" gain in performance, so don't waste the money if you don't have it. However, what about when the ratio flips back to even, i.e. not a 1:5 or 1:3 or 5:4 ratio but a 2:1 (or 1:2 however you are supposed to look at it). Have you overcome the latency caused by things not being in sync due to the non-even ratio, but still have memory running at half the speed of the processor? In other words, what you have gained is the latency improvements delivered by the faster memory but not the clock speed?

I know this question is all jumbled up, but I am assuming you guys are so aware of how this all get's put together that you probably got what I was getting at without all my blabbing.

TIA,

TIUK-3000

http://www.overclock.net/amd-gener [...] guide.html

http://www.tomshardware.com/review [...] ,1720.html

You should find most of your answers if not all of them in these articles.
I know they are about overclocking but I am sure you will be enlightend.

So you go out and get some DDR3-1333. The memory is running at 667MHz, while the FSB is running at 333MHz. The memory is running twice as fast as the FSB, do you think thats good? The FSB can only transfer so fast, whats the point of having memory that can go twice as fast? Its going to slow down as soon as it hits the FSB. Toss in the increase in latencies, and its a horrible idea to get DDR3.

DDR2-800MHz really is the best. Set your memory to 1:1, up the FSB to 400MHz, and put the CPU multiplier as high as you can. Your 800MHz memory will be running at 800MHz, everything will be in sync, and your CPU will be nicely overclocked. (not to its max, but overclocked.) Put all the money you saved by not buying DDR2-1066/DDR3 and a DDR3 motherboard into a nice video card and/or monitor. Enjoy playing your video games.

The E8500 has that big 6MB L2 cache right on the CPU die.

All 1333Mhz CPUs have a theoretical FSB bandwidth of 10666 MB/s. Two cores or four cores-its all 10666 MB/s.
When you look at the actual performance throughput its quite a bit less ~ 6975 MB/s for the E8500.
1333 FSB E8500 6975 MB/s (as benchmarked with SiSoftware Sandra Lite Memory Performance)
http://www.bit-tech.net/hardware/2 [...] nd_e8200/3

FSB bandwidth is a function of the FSB speed. More speed=more bandwidth.
333Mhz (1333 FSB) 10666 MB/s
400Mhz (1600 FSB) 12800 MB/s
450Mhz (1800 FSB) 14400 MB/s

To find optimum RAM? How to match the CPU bandwidth with RAM bandwidth?
Have you ever wondered why RAM has TWO names?
As in DDR2 667 (PC2 5400), DDR2 800 (PC2 6400), DDR2 1066 (PC2 8500) and DDR3 1333 (PC3 10666)?
The DDR part indicates data transfers per second (in millions); 677 million data transfers, etc.
The PC part indicates single channel bandwidth - as in PC2 5400=5400 Mb/s, PC2 6400=6400 Mb/s
Or PC3 10666 DDR3 1333? PC3 12800 DDR3 1600? Notice anything familiar? 1333/10666 & 1600/12800?

So what happens if you don't use optimum RAM? Not too much changes.
The actual performance is very close - only a few percentage points difference.
A E8500 with DDR2 1333 PC2 10666 RAM running at 3.16Ghz will get beat by a 3.3Ghz e8500 with DDR2 800 PC2 6400 RAM.

2GB vs 4GB vs 8GB; The RAM's primary purpose is to to keep the L2 cache filled with data and instructions (as the L2 cache keeps the L1 cache filled with data/instructions for actual execution tasks).
That's one reason why doubling your RAM from 2GB to 4GB and again to 8GB doesnt double your actual performance.
Again a faster clocked CPU (3.3Ghz) with 2GB will almost always beat a slower clocked CPU (3.16Ghz) with 4GB (or 8GB for that matter).

Bottom line - CPU frequency changes (OC'ing) have a far larger impact on performance than FSB bandwidth, RAM speed or amount of RAM installed. Your system's RAM just needs to be fast enough to keep that 6MB cache topped off.
----------

^so are you saying we should all use ddr3 RAM with 1333/1600 mhz FSB cpu's?

also, with the whole PC/PC2/PC3... i thought it just meant the different generations of RAM's... PC for DDR, PC2 for DDR2, PC3 for DDR3... you said its all about the channels. so why aren't there any DDR modules that have PC2 written on them? are you saying that all the DDR3 modules that say PC3 are in "triple channel"? i didnt even think such a thing existed.

Only if you are using a DDR3 motherboard.

>.> obviously...wat do u think im gonna try and put ddr3's in a ddr2 board...

V3nom i think me and you think the same. his comment have alot to desire.

we all know higher speed= more bandwidth but the memory doesnt ultilise all the FSB bandwidth because some are distributed to SATA,PCI-E,USB etc. so doesnt matter how fast you run the ram its never going to match the full bandwidth of the FSB.

and memory will go as fast as the FSB allows it to. a FSB800 CPU can take DDR3 memory its the same as 1333 and 1600FSB processor.

the PC2-xxxx numbers is the THEORACTICAL speed the ram can achieve when running at specific speed like DDR2-800=PC2 6400 means when the memeory running at 800mhz will have 6400mb/s bandwidth, thats what it can achive at rated speed give there are no limit factors such as system bandwidth, size of bus(64bit/128bit). according to this only half of the FSB bandwidth(12800MB/s given by you) is given to the memory subsystem when running in sync mode at FSB400. and the PC1/2/3 doesnt mean single or dual or triple channel read speed, it simply have the same meaning as DDR1/2/3.

DDR/DDR2/DDR3 & PC/PC2/PC3 does indicate the generation & design spec of the RAM - it has nothing to do with channels.
The number of memory channels is determined by the design of the memory controller which is reflected on the motherboard. AMD (memory controller on the CPU die) and Intel (memory controller on the Northbridge chip) are currently using dual channel memory architecture.
It looks like the Bloomfield version of Nehalem will use a triple channel memory controller (on the CPU die) while the Lynnfield version of Nehalem will use a dual channel memory controller (also on the CPU die) but both will use DDR3 memory.

The data following the DDx & PCx is pretty handy - speed and bandwidth. But the bandwidth rating is for the single stick in single channel. Two sticks in a dual channel motherboard would have double the bandwidth; as in two DDRx 800 PCx 6400MB/s + PCx 6400MB/s having a combined bandwidth of 12800 MB/s.
That theoretical bandwidth matches the theoretical bandwidth of a 400FSB E8500.
The actual bandwidth performance of the CPU is quite a bit less than the theoretical mark which is why there is no great penalty using PCx 5300 or no great advantage using PCx 8500 RAM.

FSB bandwidth is more limited by the CPU's quad pumped bus to the Northbridge than by RAM bandwidth to the Northbridge.

C2D/C2Q CPUs are pretty much indifferent to the RAM part of the motherboard because of the cache prefetch and large L2 cache. You get a much better performance improvement increasing the clock frequency of the CPU than by choosing RAM with higher speed and bandwidth.

With the pretty hefty increase in the Nehalem CPUs bandwidth the technical improvements in DDR3 will become more of a factor.

yes, iluvgill already said that...look

notice the DOESNT in his comment...

(Sorry for the delay on the comment...)

1) When you say RAM must be fast enough to fill L2 cache does it mean particularlly for 333MHz bus, DDR2-667 is always as fast as DDR2-800 or DDR2-1066? (ie, no additional real performance achieved without OC).

2) It was not clear to me but being synchronous (1:1 ratio FSB:RAM) is "better" than being asynchronous? Does sync brings any stability or performance improvements to the system?

3) When considering OC from 333MHz to 400MHz system clock/bus, which is a real better RAM option: DDR2-667 CL4 or DDR2-800 CL5 or DDR2-800 CL4?

Thanks.

1)the ram read/write speed would be as fast as the fsb goes.

2)for INTEL!!!!! current system which use FSB as a communication link between the CPU and RAM. so higher ram speed couple with lower FSB will not bring big read/write improvement but only lower latency. but since the latency difference 10ns will not affect the system's response time. so might as well run it in SYNC/1:1 so there is no extra "UNNECCESSARY" stress on the ram when running at high speed and voltage.

3)as have said above in your question 2.its always better to run it in sync which is DDR2-800. and if the ram is good enough lower latency CL4 will be better and faster then CL5 as it will give a lower latency but read/write speed will roughly stay the same but the copy speed will increase slightly dues to the lower latency, which mean the ram can copy more data in the same give period.