Tight Timings vs High Clock Frequencies

[Ozzie]

Should you go for memory that runs with tight timings and average frequency, or for relaxed timings at high clock frequencies? In fact, you might ask whether the whole timings vs. clock speed discussion affects performance at all?

Read this article at tomshardware

[Protoplasm72]

Tastes Great ! ...... Less filling!!!


Tastes Great !!!! ...... Less filling!!!!!!!


Tastes Great !!!!!!! ......

[Sandog]

Mushkin/Crucial have some new DDR2 modules that go 3-3-3-10 and use new Elpida memory chips. I am waiting for Micron to produce their new DDR2 memory chips before I invest in a set of 1GB DDR2 sticks. It might be a few more months until Corsair has some new 1GB sticks out which is the brand I prefer.

[SpeedDMN]

I wont try to sound like I know much since overclocking and I have a nodding acquaintance at best, but I remember reading an article describing how little memory plays a role in total system performance considering you already run a low latency dual channel setup. In a nut shell, CPU speed was far and away the critical factor in getting that high 3dmark05 score or whathaveyou. Feel free to tell me I'm wrong. I still have a lot to learn in that area

[FearFactory]

Tight memory timings will give you higher scores on memory benchmarks like Sandra but the overall performance i really dont know if there's a diference.

[SpeedDMN]

I guess maybe the real question is how do you set memory timings to accomodate the higher CPU frequencies.

[specmike]

For most of the things I have ever tried or tested, the only thing LL type memory does is provide a higher memory bandwidth score and it seems to make game play a bit "smoother". From what I have read gaming has lots of dynamic or random reads and therefore the lower latencies smooth out your gaming experience. I can tell a difference in some games with my current systemversus my older systems but, that is not apples v apples. I have never actually sat there and tried to play comparing latency settings. Seems like a waste of time. Of course, running synch vs asynch can play into all of this too but that's an entirely seperate variable that Tom did not test. As they noted, not worth the time really.

Mike.

[Acid8000]

Are things similar for Intel/DDR2 platforms?

[shaihulud]

You cant really place a value on what frequency or timing is in the matter of importance unless a careful examination of the architecture is done. This will obviously mean that it will differ. However, there are some serious errors with articles like this mainly due to cache.

Without examining the cache structure and coherency, at least, you will never comprehend what the chip is doing with the bits and accesses to memory. Most is prefetched and hidden. Disabled the L2 in the Athlon is kind of ignorant unless it is a cache miss and not in the victim buffer (q.v. exclusive). Then the L2 might matter and this can vary due to a few circumstances.

In general you want frequency and time to “equal in the inverse” with each other. Things are changing though due to different memory management units, cache architectures, core logic arbitration methodologies. To test you can't just sample core logic X and expect all to follow the same example. But in general lower frequencies and higher timings will always result in a lesser performance to some degree. End of story, tired of articles like this. They never target what is important. This has been covered since lower CAS RAM was economically viable for consumers.

[h20-ski]

First it depends on what platform you are on whether AMD or Intel, then what cpu you are using. ie A64 and Dothan tends to give better synthetic number with lower latency, while any P4 will see better results with a higher FSB than lower latency. Ideally you would want both, but unfortunately there is no ram that does not scale with frequency. The nice thing about Xeons is the unlocked multiplier, as found out running my new 3.2GHz 533MHz FSB Xeons, benefit more from being run at 16*200, than 24*133.

[DeusExMachina]

16*200 would make better sense, because 533mhz procs are quad pumpers and respond favorably to ratios that fit the board FSB speeds (being that it can handle 800mhz FSB). 24*133 would work better for a 533mhz board as that's native.

DXM