Keep ram at 800 with 3-3-3 timings, or 1000+ with looser timings?

[Jumpem]

What gives a better benefit, tighter timings or more speed?

 

[BitByBit]

Latency (/ns) = CAS / Frequency

3 / 400 (DDR2 800) = 7.5ns
4 / 500 (DDR2 1000) = 8.0ns

The additional bandwidth afforded by the latter configuration would likely not be enough to offset the (slight) latency advantage of the former.

Your best bet is to run some benchmarks and determine the best configuration for yourself.

 

[JAG87]

Originally posted by: BitByBit
Latency (/ns) = CAS / Frequency

3 / 400 (DDR2 800) = 7.5ns
4 / 500 (DDR2 1000) = 8.0ns

The additional bandwidth afforded by the latter configuration would likely not be enough to offset the (slight) latency advantage of the former.

Your best bet is to run some benchmarks and determine the best configuration for yourself.

true in an ideal situation, but its not like that..

intel benefits from more bandwith. period. the more mhz your ram has, the more performance you will get.

on AMD, the story is different because of the on die memory controller being so much better than intels northbridge memory controller. low latencies benefit more on AMD.

overall, the difference is totally negligible between 800 mhz CL3 and 1000 mhz CL4 for example. so dont kill yourself over this.

 

[n7]

If you can get 1000 4-4-4, 1000 will win in basically every real world sitation.

I even found 1063 5-5-5 beat 972 4-4-4.

FYI, synthetic benches will often favor the tighter timings, but SuperPI does not mirror real world performance.

 

[Jumpem]

Ok, thanks guys!

 

[cmdrdredd]

Higher frequency for Intel = more performance.

 

[lopri]

Originally posted by: n7
If you can get 1000 4-4-4, 1000 will win in basically every real world sitation.

I even found 1063 5-5-5 beat 972 4-4-4.

FYI, synthetic benches will often favor the tighter timings, but SuperPI does not mirror real world performance.

I think you got it backward?

Synthetic benches will often favor higher bandwidth (especially Super PI/Sandra), but in real world 800MHz/3-3-3 should be equal or better than 1000MHz/4-4-4. What's critical here (and not many reviews deal with this) is that not only the big 4 (tCL-tRCD-tRP-tRAS) but also most other sub-timings can be tightened better @800MHz than @1000MHz.

With the introduction of NF680i chipset, there is another extremely important timing came to Intel: Command Rate. Unfortunately it is already hard to achieve 1T @800MHz, and it'll be near impossible @1000MHz+. Regardless, from my testing on this chipset, 800MHz/3-3-3-1T beat everything including 1100MHz/4-4-4-2T.

P.S. Before I click the reply button, I realized what I said above was based on overclocked configuration of 400FSB, where DDR2-800 will make 1:1 with the FSB. I'd assume @266FSB, 1000MHz/4-4-4 would be slightly faster than 800MHz/3-3-3. (both async)

 

[n7]

I know 1T is huge, but 1T is only on 680i, so i didn't bother mentioning it, since it's not applicable here (OP has a P5B-D).

Also, i highly doubt DDR2-800 3-3-3 is faster than DDR2-1000 4-4-4 in anything but SuperPI or other synthetics.
Even if it's 1:1 vs. 4:5 on an overclocked system.

Run 3DMark01 @ 9x400 1:1 3-3-3 vs. 4:5 4-4-4 (w/o 1T involved) & tell me which config wins.
IMO, 01 is a very decent indicator for most games performance.

 

[lopri]

Well.. I know in Super PI and most other synthetics, 1000MHz/4-4-4 will win over 800MHz/3-3-3. But I will try 3DMark01 later on.

 

[littlezipp]

I've been running benchmarks to answer your exact question.
I have not had a conclusive answer yet, but I recommend moving your multiplier, FSB and ram timings around until you find the right combination of votage, heat and performance.

 

[Idontcare]

The first responder to this thread hit the nail on the head. 800 at CAS3 has essentially the same latency as 1000 at CAS4

In fact 800*4/3 = 1066...so CAS3 at 800 is the same as CAS4 at 1066 in that both cases have the same absolute latency in time. If your application benefits from CAS3 latency at 800 then it should see the same (not worse is my point) performance at CAS4 and 1066.

The benefit then comes from the bandwidth of 1066 over 800. For those applications which benefit from bandwidth you will want 1066 CAS4 over 800 CAS3.

Probably still not clear to many of you, but I gave it a shot anyway.

This applies to 1T vs. 2T as well. 800 with 1T would be same as a theoretical stick of DDR2 at 1600 and 2T, same effective command rate but 2X the bandwidth.

 

[littlezipp]

Originally posted by: Idontcare
The first responder to this thread hit the nail on the head. 800 at CAS3 has essentially the same latency as 1000 at CAS4

In fact 800*4/3 = 1066...so CAS3 at 800 is the same as CAS4 at 1066 in that both cases have the same absolute latency in time. If your application benefits from CAS3 latency at 800 then it should see the same (not worse is my point) performance at CAS4 and 1066.

The benefit then comes from the bandwidth of 1066 over 800. For those applications which benefit from bandwidth you will want 1066 CAS4 over 800 CAS3.

Probably still not clear to many of you, but I gave it a shot anyway.

This applies to 1T vs. 2T as well. 800 with 1T would be same as a theoretical stick of DDR2 at 1600 and 2T, same effective command rate but 2X the bandwidth.

Now would it be better to run the fsb at 400 and the ram at 1000 via dividers?
My ram can go that high (higher even - 8500C5), but my motherboard maxes out at 420ish.

 

[Idontcare]

littlezip,

I can't tell you for certain without gaining more experience myself, my understanding is that the asynchronous dividers add additional latency to the memory transactions.

Going by ancient recollection here from back in the day, maybe things have improved with this latest generation of chipsets.

The answer to your question will depend on your application's tradeoff between latency setback versus bandwidth increase when going to asynchronous mode. Note by asynchronous I mean something other than the standard even integers used today, 1:2 = ok, but 4:5 is going to mean extra latency adder by the MCH (to recollection).

 

[cmdrdredd]

It also depends on your strap setting. If you're running 400Mhz on a P5B you're on the 1066 strap with tight timings so you want every extra Mhz out of your memory at this setting. If you're below that or above it you want to try to get tighter timings since you're on the 1333 strap and you don't have the tighter timings in the strap.

so for a P5B 400Mhz FSB and DDR2-1000 5-4-4-8 beats DDR2-800 3-3-3-4 in my system.

 

[Jumpem]

Originally posted by: cmdrdredd
It also depends on your strap setting. If you're running 400Mhz on a P5B you're on the 1066 strap with tight timings so you want every extra Mhz out of your memory at this setting. If you're below that or above it you want to try to get tighter timings since you're on the 1333 strap and you don't have the tighter timings in the strap.

so for a P5B 400Mhz FSB and DDR2-1000 5-4-4-8 beats DDR2-800 3-3-3-4 in my system.

I'm at 401 to get to the 1333 strap. Not sure what that means exactly, but its what people have said to do.

 

[cmdrdredd]

Originally posted by: Jumpem

Originally posted by: cmdrdredd
It also depends on your strap setting. If you're running 400Mhz on a P5B you're on the 1066 strap with tight timings so you want every extra Mhz out of your memory at this setting. If you're below that or above it you want to try to get tighter timings since you're on the 1333 strap and you don't have the tighter timings in the strap.

so for a P5B 400Mhz FSB and DDR2-1000 5-4-4-8 beats DDR2-800 3-3-3-4 in my system.

I'm at 401 to get to the 1333 strap. Not sure what that means exactly, but its what people have said to do.

it loosens the timings on the Northbridge alot and gives better stability for alot of people. Your memory performance goes down though as a result.

 

[Jumpem]

Originally posted by: cmdrdredd
it loosens the timings on the Northbridge alot and gives better stability for alot of people. Your memory performance goes down though as a result.

So am I better off at 9x400, than 9x401 if it stable?

 

[cmdrdredd]

Originally posted by: Jumpem

Originally posted by: cmdrdredd
it loosens the timings on the Northbridge alot and gives better stability for alot of people. Your memory performance goes down though as a result.

So am I better off at 9x400, than 9x401 if it stable?

if 9x400 is stable you will see better performance

 

[lopri]

I tried to compare 800MHz/3-3-3 with 1000MHz/4-4-4 on my board (680i) using 3DMark01, and my conclusion is - it's impossible! The scores vary from a run to another by larger margins than the differences resulted from memory configuration. After a couple of runs, I decided to test only Game 3 (Lobby) to save time. And for whatever reason, 2nd run is slower than 1st run, and 3rd run is slower than 2nd run. After the 3rd run the scores somewhat stabilize but it's far from being consistent.. I'd need a test that's more repeatable than '01. Maybe I should try Doom 3?

 

[lopri]

Originally posted by: BitByBit
Latency (/ns) = CAS / Frequency

3 / 400 (DDR2 800) = 7.5ns
4 / 500 (DDR2 1000) = 8.0ns

The additional bandwidth afforded by the latter configuration would likely not be enough to offset the (slight) latency advantage of the former.

Your best bet is to run some benchmarks and determine the best configuration for yourself.

I think we've graduated from the theoretical latency values? FSB, dividers, and chipset frequency/latency are as big as memory frequency/timing itself these days. I totally agree with the latter part of your post, though.

Edit: As a matter of fact, looking at how memory works in 680i makes all the memory tests performed in the past look as if they were chipset/motherboard tests, rather than memory tests. They are still relevant in that better performing memory on 975X will likely better perform on 680i, but I am seeing so many erroneous conclusions (well, to be considered as pure memory test) in many memory reviews.

 

[cmdrdredd]

I just know that from my own testing DDR2-800 at 3-3-3-4 is slower in all apps I can test than DDR2-1000 5-4-4-8

 

[lopri]

I was indeed wrong with my understanding. Ran Doom 3 timedemo and the results were roughly as following:

800MHz/3-3-3: 193 FPS
1000MHz/4-4-4: 194 FPS
1000MHz/5-5-5: 192 FPS

While the differences are tiny, they were consistent. So on this board (680i) and likely on other Intel platform, DDR2-1000/CL4 is faster than DDR2-800/CL3. I stand corrected.

 

[Jumpem]

Originally posted by: lopri
I was indeed wrong with my understanding. Ran Doom 3 timedemo and the results were roughly as following:

800MHz/3-3-3: 193 FPS
1000MHz/4-4-4: 194 FPS
1000MHz/5-5-5: 192 FPS

While the differences are tiny, they were consistent. So on this board (680i) and likely on other Intel platform, DDR2-1000/CL4 is faster than DDR2-800/CL3. I stand corrected.

Good information, but at the same time the difference is so minute that I shouldn't spend much thought on one or the other.

 

[n7]

Nice to see my conclusions were actually relatively correct.

That being said, it is very true that the difference between bandwidth with looser timings vs. lower bandwidth + tight timings is extremely minimal.

So i'd say it's not a big deal how you choose to run your RAM

 

[betasub]

Originally posted by: n7
That being said, it is very true that the difference between bandwidth with looser timings vs. lower bandwidth + tight timings is extremely minimal.

So i'd say it's not a big deal how you choose to run your RAM

QFT. So ppl should QFW