Sign in with
Sign up | Sign in
Your question

Does Latency Matter?

Last response: in Memory
Share
May 6, 2007 6:51:57 PM

Ive been looking a lot at RAM lately, trying to find some that will fit in my PC.
Ive noticed that the price of lower latency RAM is much much higher than that of higher latency. For instense the Corsair DDR2-800 CL 4 modules cost around 200$, while the CL 3 modules cost over 350$.
So logically I assumed that the difference between CL 4 and CL 3 is huge.

However after looking at many reviews, it appears that the difference is actually less than 5%.
So from this i have to assume that latency DOES NOT matter, and i can have DDR-800 CL 10 and wouldnt be losing too much performance.
But then, whats with the much higher price for lower latency modules?

Thnx!

More about : latency matter

May 6, 2007 7:00:27 PM

So the only reason to buy DDR2-800 CL4 instead of DDR2-800 CL5 is because the CL4 modules will allow u to overclock to DDR2-1066 CL5?
May 6, 2007 7:00:40 PM

Just as with CPUs, there is not a linear relationship between performance and price.
Partly this is due to the lower yield of chips near the performance extremes (e.g. high speed/low latency), partly this is because the manufacturers know that buyers who want/need items near the performance extremes are willing to pay a disproportionate price for them.
Related resources
May 6, 2007 7:04:26 PM

Quote:
Just as with CPUs, there is not a linear relationship between performance and price.
Partly this is due to the lower yield of chips near the performance extremes (e.g. high speed/low latency), partly this is because the manufacturers know that buyers who want/need items near the performance extremes are willing to pay a disproportionate price for them.


But thats my point.
These "performance extreme" modules arent better than the ones normal ppl buy, wich is UNLIKE CPUs where there is a huge difference between 2Ghz and 3Ghz.
May 6, 2007 7:13:42 PM

Quote:
...
These "performance extreme" modules arent better than the ones normal ppl buy

Yes, they are. They should run faster and/or at lower latencies.
Quote:
..., wich is UNLIKE CPUs where there is a huge difference between 2Ghz and 3Ghz.

There's just as huge a difference in *memory* performance between DDR2-533 modules and DDR2-800 modules.
May 6, 2007 7:17:07 PM

Quote:
...
These "performance extreme" modules arent better than the ones normal ppl buy

Yes, they are. They should run faster and/or at lower latencies.

You're not listening to what im saying.. sorry to be rude.
YES, they are lower latency, but they do not have better performance because of this.
May 6, 2007 7:33:38 PM

With respect, I'm listening (reading) more closely than you are: you're not being clear because you're making unqualified blanket statements rather than precisely-qualified ones. I'm just trying to clarify things. :) 

Of course RAM run at lower latencies has better *memory* performance than RAM run at higher latencies (at the same speed). The former also normally leads to increased *system* performance as well (except perhaps at FSB:memory throughput ratios slightly below 1:1).

However, (and I'm guessing this is your point) the *degree* of system performance improvement due to the lower latencies is often so small as to not be easily noticeable during normal use. It's useful to look at real-world data and decide for yourself where your money is best spent, also taking into account the possibility of re-using RAM on a future system with a likely-higher FSB.

Here's some data for Core2Duo systems (AMD x2 AM2 systems will have different results due to their built-in memory controller): http://www.madshrimps.be/?action=getarticle&number=1&ar...
May 6, 2007 8:18:08 PM

Quote:


However, (and I'm guessing this is your point) the *degree* of system performance improvement due to the lower latencies is often so small as to not be easily noticeable during normal use. It's useful to look at real-world data and decide for yourself where your money is best spent, also taking into account the possibility of re-using RAM on a future system with a likely-higher FSB.


Alright, so if we agree that the performance difference is so small that it is not noticeable, then why do CL3 modules cost so much more than CL4 modules? People arent stupid.. they wont buy something that is expensive just for that fact.

And since im going to have my FSB at 2133Mhz or higher, i doubt ill be re-using these modules for my next build, wich will likely have a much higher FSB.
May 6, 2007 9:05:16 PM

Nice post...Madshrimps cleared up some questions I had,although there are a few more that are pretaining to ddr2.On my last build I went with corsair 5-5-5-12(actually 15)an it worked out pretty well oc'ing a gigabyte 945p-s3 with a d805.My best advice is go to reputable sites and see what they recommended for your application(dont know if your wanting to do some serious oc'ing or not) :D  gl.
May 6, 2007 9:49:45 PM

Interesting link. Don't fail to notice, however, that their bar graphs don't start at 0. There's a max differential of ~45% on the sandra tests, and about ~10% on the SuperPI tests.

It's too bad they didn't do any media encoding or gaming benchmarks.
May 6, 2007 10:20:12 PM

Quote:
...
Alright, so if we agree that the performance difference is so small that it is not noticeable, then why do CL3 modules cost so much more than CL4 modules? ....

CL3 is pushing the margins of DDR2 technology at speeds like DDR2-800, so the yield of chips able to run at that latency has got to be substantially less than the yield of CL4-capable chips.
Also, while you and I wouldn't do so, some people ARE willing to pay very high price increments for only a small performance increment. For example, a QX6700 costs $970, 80% more than the Q6600 for only a roughly 11% clock speed boost.
May 6, 2007 11:00:26 PM

Yea, it's a quick and dirty benchmark, kind of only confirms what most of us already know.

Remember, when you increase frequency there's a massive bandwidth gain that you can't get from tighter timings. High freq, loose sticks will always be faster than low freq tight.
May 7, 2007 4:03:41 AM

I've heard both sides of that theory but I usually buy the tighter latency modules,most of my machines run faster than others in real world applications :D  .
May 7, 2007 8:35:14 AM

Quote:
...
Alright, so if we agree that the performance difference is so small that it is not noticeable, then why do CL3 modules cost so much more than CL4 modules? ....

CL3 is pushing the margins of DDR2 technology at speeds like DDR2-800, so the yield of chips able to run at that latency has got to be substantially less than the yield of CL4-capable chips.
Also, while you and I wouldn't do so, some people ARE willing to pay very high price increments for only a small performance increment. For example, a QX6700 costs $970, 80% more than the Q6600 for only a roughly 11% clock speed boost.

So ppl actually pay twice as much money for a 5% performance increase?
Thats insane.

Your CPU analogy is wrong however. Because i can easily gain an 11% overclock on the Q6600, but i will never be able to get to 1066 CL4 from 667 CL4.
May 7, 2007 8:36:09 AM

Quote:
I've heard both sides of that theory but I usually buy the tighter latency modules,most of my machines run faster than others in real world applications :D  .


That graph clearly demonstrates that tighter timings are meaingless and offer no performance increase at all.

How can u possibly disspute this?
May 7, 2007 9:16:07 AM

Tighter timings DO offer a performance increase, it's just not generally as big as higher frequency. I generally try to purchase low-latency myself, simply because it tends to overclock further as it is more tightly binned. If I have to loosen it up to get the higher freq, well, so be it.
May 7, 2007 11:15:30 AM

So ppl who buy DDR2-1111 at CL4 for 600$ arent crazy because they actually want to overclock to 1350Mhz CL5?
May 7, 2007 11:41:50 AM

Did I say that? No. That's a little excessive.

However, there's nothing crazy about getting 4-4-4-12 DDR2-800 2.0V for $135 and bumping it to 1066 5-5-5-18 @ 2.2V, right?
May 7, 2007 12:52:58 PM

Quote:
Did I say that? No. That's a little excessive.

However, there's nothing crazy about getting 4-4-4-12 DDR2-800 2.0V for $135 and bumping it to 1066 5-5-5-18 @ 2.2V, right?


But then why do ppl buy 1066 4-4-4-12?
May 7, 2007 2:47:50 PM

To get 1333 @ 5-5-5-15? Why do people buy name-brand peanut butter? Why do people by name-brand medication? Why let it bother you?
May 8, 2007 12:04:41 AM

Just settle this debate with numbers....

Say you have a 800 CL5 and a 800 CL4....

Now the true clock frequency is 400 MHz, so one clock cycle is 1/(400 MHz) = 2.5ns.

So 800 CL5:

2.5ns x 5 clock cycles = 12.5 ns

800 CL4:

2.5ns x 4 clock cycles = 10 ns.

Even though it is a little more complex than this, when comparing the two Cas Latencies, you can see that CL4 is theoretically faster by a small margin.... So if you want to go faster CL4 IS FASTER THAN CL5.
May 8, 2007 2:16:32 AM

That's never been a debate. The numbers speak for themselves. Dropping one latency level usually nets a 1-5% increase, depending on the app. However, and a big however, frequency matters MUCH more than latency in the majority of applications, especially on an Intel system.
May 8, 2007 11:36:28 AM

Quote:
Just settle this debate with numbers....

Say you have a 800 CL5 and a 800 CL4....

Now the true clock frequency is 400 MHz, so one clock cycle is 1/(400 MHz) = 2.5ns.

So 800 CL5:

2.5ns x 5 clock cycles = 12.5 ns

800 CL4:

2.5ns x 4 clock cycles = 10 ns.

Even though it is a little more complex than this, when comparing the two Cas Latencies, you can see that CL4 is theoretically faster by a small margin.... So if you want to go faster CL4 IS FASTER THAN CL5.


Thank u for that.
But i have no idea what that means..

Yes, 5 clock cycles on 400Mhz are 12.5ns, but what does that mean?
May 8, 2007 12:37:19 PM

And after all those answers , i would suggest you buy the "slower" ram and get yourself a better CPU or GPU to actually see a difference. :) 
May 8, 2007 4:20:45 PM

Yeah, I would agree with the faster CPU over the faster RAM, but here is an overview of Cas Latency....

When the computer retrieves information, it takes it from the page files in 4 word bursts....where if there is a page hit (kind of like if the information is already in the RAM), only the Cas Latency and the frequency matter...

Now according to Intel, a clean page hit is about 30% of the time...

So, about 30% of the time, your page hit will be retrieved in 10 ns instead of 12.5 ns, assuming your memory is your only bottleneck on the system (usually not true though). :) 

Even though I didn't explain this very well, I just didn't feel like typing an essay out... Sorry if I couldn't explain it better....
May 8, 2007 9:31:11 PM

Quote:


When the computer retrieves information, it takes it from the page files


Lost me.
Page files?
Quote:

in 4 word bursts


No idea what that means.
Quote:

....where if there is a page hit


A what?

Quote:

only the Cas Latency and the frequency matter...


What matters other than the frequency and latency?



Quote:

So, about 30% of the time, your page hit will be retrieved in 10 ns instead of 12.5 ns, assuming your memory is your only bottleneck on the system (usually not true though). :) 


So those 2.5ns actually matters? I mean i still have no idea what 10 and 12.5ns actually means.. what is it?
May 8, 2007 9:31:50 PM

Quote:
And after all those answers , i would suggest you buy the "slower" ram and get yourself a better CPU or GPU to actually see a difference. :) 


I already have an 8800 GTX and an E6600.. what else could i possible upgrade?
May 8, 2007 9:46:04 PM

Track,

12.5ns to 10 ns is about a 20% gain, if the only thing bottlenecking your system is memory....

But you do not see a 20% performance increase because of other factors, but I was just trying to explain the theoretical advantage...

It's all up to you if you want to go with the value ram, or the overclocked ram....
May 8, 2007 11:22:31 PM

He's just getting a bit (confusingly) into the mechanics of how memory is managed at a system level. Don't worry about it.
May 9, 2007 12:21:40 AM

Haha, yeah... I could barely read and understand the post myself.... Sorry for the confusion.
May 9, 2007 7:21:56 AM

Quote:
And after all those answers , i would suggest you buy the "slower" ram and get yourself a better CPU or GPU to actually see a difference. :) 


I already have an 8800 GTX and an E6600.. what else could i possible upgrade?

ok then , i cant see your problem.You wont see any real improvement by better ram.To give you an example , i tried mine at 1000 MHz and got nothing more than 800 Mhz(at least something you can notice).Unless you want to be happy by looking at the bandwidth of the higher freq.Having 2Gb or even 4Gb of ram at the cost of a quick 1GB is better.
May 9, 2007 3:02:57 PM

Quote:
Track,

12.5ns to 10 ns is about a 20% gain, if the only thing bottlenecking your system is memory....

But you do not see a 20% performance increase because of other factors, but I was just trying to explain the theoretical advantage...

It's all up to you if you want to go with the value ram, or the overclocked ram....


That dosent answer my question at all..

Yeh, 12.5ns is 25% slower than 10ns, but what does that mean? I mean 12.5ns EVERY TIME the RAM is accessed, or 12.5ns every second wich would be next to nothing?

Also, im confused about the 12.5ns thing..
I mean, its not 400 clock cycles, its 400,000,000 clock cycles. So it would be 12.5 micro seconds or something..

And even if it was 400 clock cycles, that means that 500 clock cycles (500Mhz) CL5 would result in 10ns.. so the higher the freaquency the higher the bandwidth but also, the lower the latency.

So the latency of 400Mhz CL4 is the exact same as the latency of 500Mhz CL5.
May 9, 2007 3:08:16 PM

Quote:
And after all those answers , i would suggest you buy the "slower" ram and get yourself a better CPU or GPU to actually see a difference. :) 


I already have an 8800 GTX and an E6600.. what else could i possible upgrade?

ok then , i cant see your problem.You wont see any real improvement by better ram.To give you an example , i tried mine at 1000 MHz and got nothing more than 800 Mhz(at least something you can notice).Unless you want to be happy by looking at the bandwidth of the higher freq.Having 2Gb or even 4Gb of ram at the cost of a quick 1GB is better.

So ur saying i should buy a lot of slow and poor performing RAM?
May 9, 2007 3:10:06 PM

Quote:
He's just getting a bit (confusingly) into the mechanics of how memory is managed at a system level. Don't worry about it.


Why shouldnt i worry about it?

I know everything else about hardware, might aswell know about RAM.
May 9, 2007 4:00:40 PM

I meant "don't worry about that specific post". Chill a bit - you angst too much. :) 
Sounds like you want to learn how RAM works down at the wiring level. There's an excellent multimedia presentation on this by Corsair that's linked to in the Forum sticky FAQ (right at the end).
May 9, 2007 4:29:49 PM

Mondo,

Thanks for the direction to the Corsair post.... man, they can explain it a lot better.... I must have not seen it before.
a b } Memory
May 9, 2007 8:02:36 PM

I have been following this thread for a couple of days. Got curious. Loaded SiSoft Sandra 2007 Lite into the new box.

Going from the "stock" 5-5-5-15-2T timings to 3-3-3-7-1T, I pick up about a 6% increase in memory bandwidth.By itself, I do not think the increase is worth the extra cost. However, I shoud be able to drop a Penryn CPU in with only a BIOS upgrade.
May 9, 2007 8:38:24 PM

Quote:
I have been following this thread for a couple of days. Got curious. Loaded SiSoft Sandra 2007 Lite into the new box.

Going from the "stock" 5-5-5-15-2T timings to 3-3-3-7-1T, I pick up about a 6% increase in memory bandwidth.By itself, I do not think the increase is worth the extra cost. However, I shoud be able to drop a Penryn CPU in with only a BIOS upgrade.


You have DDR2-1000 at CL3?
May 9, 2007 8:40:36 PM

Quote:
I meant "don't worry about that specific post". Chill a bit - you angst too much. :) 
Sounds like you want to learn how RAM works down at the wiring level. There's an excellent multimedia presentation on this by Corsair that's linked to in the Forum sticky FAQ (right at the end).


But my theory that 400Mhz CL4 and 500Mhz CL5 have the EXACT same latency of 10ns is still true.
May 9, 2007 8:48:06 PM

Quote:
So ur saying i should buy a lot of slow and poor performing RAM?


If you read the answers you got from all people in here you should be able to understand that "slow and poor performing" is as good as "expensive and fast" in most situations.As i said if you have the cash go buy whatever you want,you dont have to bother that much.
May 9, 2007 8:49:41 PM

Quote:
So ur saying i should buy a lot of slow and poor performing RAM?


If you read the answers you got from all people in here you should be able to understand that "slow and poor performing" is as good as "expensive and fast" in most situations.As i said if you have the cash go buy whatever you want,you dont have to bother that much.

Good answer.
Can u tell me why i want that extra 6% so badly?
May 9, 2007 9:07:18 PM

Man, jsc you have your memory at CL3 with 1T @ 1000MHz... impressive.

Ive heard that 1T is about 5-10% faster than 2T, wonder if that is giving you the biggest performance increase....
May 9, 2007 9:48:16 PM

Quote:
Man, jsc you have your memory at CL3 with 1T @ 1000MHz... impressive.

Ive heard that 1T is about 5-10% faster than 2T, wonder if that is giving you the biggest performance increase....


CL3 @ 1000Mhz is impossible. 1T is just crazy.
May 9, 2007 11:01:32 PM

Quote:
I meant "don't worry about that specific post". Chill a bit - you angst too much. :) 
Sounds like you want to learn how RAM works down at the wiring level. There's an excellent multimedia presentation on this by Corsair that's linked to in the Forum sticky FAQ (right at the end).


But my theory that 400Mhz CL4 and 500Mhz CL5 have the EXACT same latency of 10ns is still true.

Right....but one has 100,000,000 cycles more per sec. The CL5 not matter at all as you still have a large preformance gain that more than offsets the 1 clock.

On older DDR1 systems a 30% boost is very noticable (useing DDR-520 on a DDR-400 system 1:1).
a b } Memory
May 10, 2007 6:28:35 AM

Quote:
I have been following this thread for a couple of days. Got curious. Loaded SiSoft Sandra 2007 Lite into the new box.

Going from the "stock" 5-5-5-15-2T timings to 3-3-3-7-1T, I pick up about a 6% increase in memory bandwidth.By itself, I do not think the increase is worth the extra cost. However, I shoud be able to drop a Penryn CPU in with only a BIOS upgrade.


You have DDR2-1000 at CL3?

I wish. What I have is DDR2-1000 memory running at DDR2-733 speed at CL3. I have an E6600 running at 3.3 GHz (367 MHz X 9). So memory clock is 733 MHz. My apologies to all for not being more clear.

When I bought the parts for the new box in December, I tried to future proof as much as possible. My E6600 will run at 3.6 GHz (FSB 400 MHz) only with stability and thermal problems. The motherboard FSB will reach 460 MHz. Memory will run at an FSB of 500 Mhz.

Memory Master asked about the difference between -1T and -2T timings. I was curious - I spend a lot of time being curious - so I checked. Going from -2T to -1T accounts for a little less than half of the 6%.

I will concede that, outside of benchmarks, a 6% increase is not noticeable. I will also admit that that 6% cost me about $60 after rebates. As CPU's get faster, demand for fast memory will increase. I do not expect the price of really fast memory to drop significantly in the near future. And upgrading to Penryn without needing to upgrade memory is a really attractive idea.

Updated signature. I think DDR2-number is easier than PC-number to understand.
May 10, 2007 12:31:58 PM

Quote:


Updated signature. I think DDR2-number is easier than PC-number to understand.


still says 1000 @ cl3
a b } Memory
May 10, 2007 3:37:50 PM

Quote:


still says 1000 @ cl3


Ah. I was focused on the body of the article. :oops: 

OK.

There. :) 
May 10, 2007 4:41:31 PM

You have to specifiy the speed..
May 10, 2007 6:13:41 PM

No, no one can tell you WHY you'd want the extra 6% so badly, only you can answer that. Stop being a pain and go read about Colum Access Strobe or "CAS" instead of trying to form a counterpoint without the technical understandings of what your arguing.
!