DDR3 latency

[hat]

Does it seem to anyone else that DDR3 latency is too high? I was just looking at some Crosair XMS3 memory, which is supposed to be high-end... it was rated for 9-9-9-24 at 1333MHz! I think 9-9-9-24 is way too high for 1333MHz, espically for memory that's supposed to be high end... I think it's even too high for value ram! What's the deal with the high latency DDR3?

 

[Fourstaff]

You do get nice CAS 7 sticks at 1600Mhz:G.SKILL Ripjaws Series 4GB (2 x 2GB) 240-Pin DDR3 ...

 

[LAN_deRf_HA]

Well corsair often is way more conservative than they need to be with their timings, but in general 1333 or lower ram is considered low end, and I've noticed not many manufacturers bother trying to make decent timings for low end ddr3. Interestingly a lot of cheap 1.5v kits can hit cas 7 or 8 even when overclocked to @1600mhz with a voltage bump.

On a side note, I dropped my ram down from 1600 to 1200mhz and lowered the timings. I swear everything is smoother despite losing performance in every benchmark.

 

[Bot]

yes when ddr3 started out, 1333 with cas9 was the norm
cas7 and cas8 are the norm now for 1600 and up

the xms3 set you saw must be Android older set, there are also xms3 set with lower cas And/or faster

 

[hat]

Yeah, I just dropped from 1000MHz 5-5-5-15 to 900MHz 4-4-4-12... it does feel better.

I know 1333 is low for DDR3, hell it goes over 2000, but still, Ballistix and XMS are names that have always been associated with high-end... so I would figure if I saw Ballistix on 1333, it should have timings like 5-5-5-15 or at least 6-6-6-18...

 

[Mussels]

Yeah, I just dropped from 1000MHz 5-5-5-15 to 900MHz 4-4-4-12... it does feel better.

I know 1333 is low for DDR3, hell it goes over 2000, but still, Ballistix and XMS are names that have always been associated with high-end... so I would figure if I saw Ballistix on 1333, it should have timings like 5-5-5-15 or at least 6-6-6-18...

all you're paying for is the name with the 'high end' stuff. i got some 1600Mhz CL7, but if it wasnt second hand it'd be way too costly for what you get.


edit: oh and dont forget that with timings, they get looser as the Mhz goes higher. DDR1 capped out at 2-2-2-5 remember?

 

[LAN_deRf_HA]

I don't get why dropping cas down one notch makes such a noticeable smoothness change but makes as little as 1% benchmark difference.

 

[hat]

I could have sworn I saw CL3 DDR1 kits... I wasn't very much in the tech scene when DDR1 was the shit, DDR2 was current when I came about...

 

[Mussels]

I could have sworn I saw CL3 DDR1 kits... I wasn't very much in the tech scene when DDR1 was the shit, DDR2 was current when I came about...

DDR1 went from 2-2-2-5 to 3-4-4-12, iirc.

 

[LAN_deRf_HA]

I remember the kit I had could go down to 2.5. Always thought that was weird choice. That kit was value corsair ram, still scored 5.9 in vista.

 

[Mr McC]

At the frequencies we're discussing, I don't think latencies have that much impact, unless benchmarking crowns are important.

 

[hat]

Latency always matters... insane bandwidth is no good if it takes forever to access it.

 

[Hayder_Master]

yeah right mine is kingmax 2200 but with 10-10-10-28, maybe some guys see it's cool but for me i see it's high since my last crucial ddr2 can run 1200 at 6-5-6-15 at 1.8v only

 

[hat]

Hmm... do you have Everest Ultimate? Would you mind running the benchmark tool and posting the results? 2200 is very fast but I'm curious to see what 10-10-10-28 does to it... and what voltage are you running that at? here's mine

 

[hat]

Ganged mode seems to give better performance

 

[Mr McC]

Latency always matters... insane bandwidth is no good if it takes forever to access it.

I read a pretty thorough test on Tom's Hardware where they reached the conclusion that latency and frequency did not have that much of an impact past a certain point. I can't vouch for the testing methods or conclusion, but it convinced me:

In all cases, we’ve seen significant performance differences when looking at the synthetic or low-level benchmarks. Memory bandwidth does increase considerably if you speed up the memory transfer rate, and tightening timings also improves performance by cutting latencies. However, only a marginal fraction of these benefits actually arrive at the application level.


I wasn't arguing that latency does not matter, but rather that, past a certain point, it doesn't matter that much, unless you're going for benchmarking crowns.

http://www.tomshardware.com/reviews/core-i7-870-1156,2482-9.html

 

[Mussels]

Ganged mode seems to give better performance
http://i47.tinypic.com/xognqh.jpg

ganged is faster for a single app/thread, unganged is for multithreaded/multiple apps

 

[Dent1]

Crosair XMS3 memory, which is supposed to be high-end... it was rated for 9-9-9-24 at 1333MHz! I think 9-9-9-24 is way too high for 1333MHz, espically for memory that's supposed to be high end

I would figure if I saw Ballistix on 1333, it should have timings like 5-5-5-15 or at least 6-6-6-18...

The default latencies are high to ensure that its compatible with a wide range of motherboards, if the memory is too tight by default you will get people with rubbish boards complaining about lock ups and blue screens of deaths because their motherboard sucks. This is one of the reasons why most motherboards and memory set 2T timings opposed to 1T timings by default to ensure that extra compatibility

Now with these high end sticks you can optionally lower the latencies (if you are an experienced overclocker) without compromising on the memory bus. For example my G.Skills are clocked @ 1800Mhz with 7-9-9-20-11-1T @ 1.8v 24-7

 

[cadaveca]

DDR1 went from 2-2-2-5 to 3-4-4-12, iirc.

You got it. The high latency DIMMs were either Samsung TCCD, or Infineon..uh...crap...well, I still have my set of mushkin redlines for 250mhz 3-3-2-9, which had infineon ICs.

When considering latenc being important, we need to look at cpu's L3 cache. As long as we can match L3 bandwidth, Latency isn't too important.

 

[Super Sarge]

Try Mushkin red lines for DDR 3 these are sticks I use

 

[W1zzard]

you guys realize that latencies are in clocks?

so at 100 mhz a latency of 1T is equal to a latency of 10T at 1000 mhz (both is 10 ns)

 

[Mussels]

you guys realize that latencies are in clocks?

so at 100 mhz a latency of 1T is equal to a latency of 10T at 1000 mhz (both is 10 ns)

actually, i barely understand it. a guide would be good *nudge nudge*

 

[Mr McC]

actually, i barely understand it. a guide would be good *nudge nudge*

Seconded *wink wink*

 

[Delta6326]

Seconded *wink wink*

i third because im getting ready to build a new ddr3 computer i don't understand it at all im used to ddr2

 

[Mr McC]

I was googling latencies and found this interesting quote:

Looking at the table you still see a latency rise with the example of CAS 5 at 1000 is slightly faster than than CAS 6 at 1150 which is slightly faster than CAS 7 at 1333, things change though at DDR1600 where CAS 8 is as fast as cas5 at 1000 so you need to examine what is actually faster and tune your system accordingly.



[source http://www.thetechrepository.com/showthread.php?t=160]

I don't know the truth of these statements, but I think they raise a couple of interesting issues and if we answered them, we might have a better understanding of what's going on:

a) Different DDR latencies have a minimal effect up to 1600 Mhz

b) Cas 8 at 1000 is comparable to CAS 5 at 1000

If we answer the question raised in point b, even if we still don't understand the inner workings of latency/frequency interaction, we at least would have a means of comparing different timings. Then we would have to answer the implications of DDR 5 at 1000 in terms of applications and/or benchmarking.

I'd be happy if I understood that.