Originally posted by: er12345
okay now i am confused. I am not a computing student so please bear with me on this question.
Is it the clockspeed or the latencies that dermine how well the memory performs in games?
So if my CPU was at 2.4GHZ (266x9) do i get any increased RAM performance by having my ram at 400mhz rather than 266mhz?
Both. But there seems to be an ongoing "discussion" -- even fueled a bit by little ol' nobody me -- whether tight latencies with lower frequency or looser latencies with higher frequency makes a bigger difference.
Obviously, you would like to push the memory speed as high as possible, and still keep the latencies tight. But this is a three-dimensional space of variable relationships. It involves (1) the speed in Mhz, (2) the latency settings, and (3) the memory voltage.
Let me try and explain. [The reason I logged on anyway this evening was to make pronouncements about my "latest discoveries" per 1:1 and other CPU-to-RAM frequency ratios.]
If you can, you'd like to use a 1:1 CPU-to-RAM ratio. This means that if the CPU "external frequency" or "CPU-FSB" speed is 266 Mhz, then the double-data-rate frequency for memory would be 533 Mhz. DDR means that the actual memory bus speed is also 266, and 266 * 2 [DDR, remember] is 533.
For me, I explored getting some high-performance DDR2-1000 RAM and under-clocking it so that over-clocking my Q6600 to 3 Ghz meant an FSB of 1,334 Mhz and a RAM [DDR] speed of 667 Mhz. Here, you wonder, what's so darn good about taking RAM that can run at 1000, and making it run at 667?
At 667, I could set the basic latencies to 3,3,3,6 and run a 1T command-rate. At 800, I would need to loosen the latencies to at least 4,3,4,9. I might not be able to use a 1:1 CPU-to-RAM speed ratio. Or -- I might have to push the CPU voltage too high.
But at a 1:1 ratio, you are also stressing the NorthBridge core, and depending on the chipset and motherboard, you might find a barrier to pushing the clock speed further. In fact, because you really don't want to trash your expensive components, you might want to limit your voltage settings to just above the maximum recommended for the CPU, and just below (or at) the recommended maximum for the memory. So you say to yourself: "I will not cross this line for VCORE, and I won't cross that line for VDIMM."
Surprisingly, I discover that with either DDR2-800 or DDR2-1000 RAM, I need to increase voltage closer to the recommended warranty maximum to get stability with super-tight latencies in the lower speed range with a 1:1 ratio.
I also discover that running DDR2-800 near it's 800 Mhz spec at a different memory multiplier/divider -- say, 4:5 -- and at even tighter latencies than certified for that memory (but looser than those in the lower range), the voltage required for stability is lower (therefore -- safer.)
So now, with a set of DDR2-800 modules (Crucial Tracers with D9 chips) -- I found a good setting with Q6600/2.4Ghz @ 3.0 Ghz, FSB = 1,334, RAM = 667, and VDIMM/RAM-voltage at 2.15V with timings 3,3,3,8,1T. The Everest memory benchmark results put the "Read" bandwidth at 9,200+ MB/s, with "Write" and "Copy" results between 6,000 and 6,400, and overall memory latency at 54 ns.
That's with a ratio of 1:1.
With a ratio of 4:5, pushing the CPU speed a little higher to 3.16 Ghz, I have to run the RAM at DDR2-880 or 880 Mhz. But these modules are "small wonders:" they're rated at 4,4,4,12 at 800, but I can run them at 4,4,4,10 at 880 Mhz and command-rate = 1T, with the voltage at only 2.125V. The Everest benchies show about 9,650 "Read", 6,500 "Write," and maybe 6,700 MB/s "Copy" -- with latency at 50 ns.
And if I drop the speed to 870 (with a corresponding reduction of CPU speed), I can run them at 4,3,4,9,1T at the same voltage. I might get similar -- but not better -- benchmark scores.
With the 1:1 ratio, if I wanted to increase bandwidth as indicated by the synthetic benchmarks, raising the CPU speed again from 3 Ghz to between 3.15 and 3.2 Ghz, I would also have to increase the memory voltage beyond 2.15V -- or loosen the latencies. Since the monitored RAM voltage at 2.15 is closer to 2.19, and since I just burned out those DDR2-1000's at 2.2V, I don't want to go there with the DDR2-800 Tracers. And if I loosen the latencies, then the benchmark scores will drop back somewhat.
Otherwise, there's not much difference in performance between {tight latencies, slower speed} and {loose latencies, higher speed} --- even if getting the latter combination means running a ratio that is not 1:1. If I push 1:1 settings to the limit, I'm going to stress components, and I don't want to push my VCORE up beyond a certain value.
There are certain ratios that are "recommended" as "optimal," among which are 1:1 and 4:5.
Keep in mind I've only spoken to "synthetic benchmarks" here, but there MUST be SOME SORT of BASIS for COMPARISON. Ultimately, you'll only be able to tell how "real-world" performance is affected by running your games at proven, stable settings.
An Afterthought: It seems to me that the motherboard makers for this DDR2 generation of product were trying to anticipate CPU speeds which 'weren't there, yet," and the memory-makers were forging ahead with the same sort of anticipations. If you run a processor that is spec'd for 1,333 Mhz FSB speed, then -- bully for you -- you can probably push the RAM to its rated spec at a 1:1 ratio -- more easily.
Another afterthought: Your overall performance will be a balancing act between cache and memory performance, so it also involves both CPU speed and memory speed -- whatever CPU-to-RAM ratio you finally use.
)
Facebook
Twitter