RAM, 1866 is obviously better than 1600 isn't it?

[Nec_V20]

When you look at RAM for your performance PC you will see the likes of "1866", "2133" or "2400" and of course these have to be a lot faster than "1600" modules.

Well let's take a closer look at the specs of the 1866 RAM, and we will see something like "CL10". Now we look at the cheaper 1600 RAM and we see "CL9". What does this mean?

The 1866 RAM is about 16.5% faster than the 1600 RAM - so the choice is a no-brainer.

But wait, at CL10 vs CL9 the 1866 RAM is 12% slower than the 1600 RAM.

Let's do the arithmetic, (1866 / 10) / (1600 / 9) *100 = 104.625%

This means that you are paying quite a bit more for a a measly 4.5% performance gain.

As it says in the Bible (and every EULA - never mind "Warranties"), "The large print giveth and the small print taketh away".

For the purposes of overclocking, RAM "speed" is the least performance enhancing factor. I would defy anyone to run a system with settings at 1600 and then at 2400 and tell me they notice any difference. If you save the money for the "high performance" RAM and stick that into a processor which has 2MB more cache then you will get a hell of a lot more of a performance boost for your money.

Let's take an example, 2*8GB 2400 RAM will cost you about £70 ($105) more than 2*8GB 1600 RAM, whereas you could save that £70 ($105) and buy an i7-4770k Haswell processor instead of an i5-4670k Haswell processor.

Now given the same motherboard which do you think will be the higher performance machine for the price? A system with an i7-4770k and 1600 RAM or the system with an i5-4670k and 2400 RAM?

The whole RAM "speed" hype is essentially a con game. On any benchmark you would like to use which is not RAM specific (and even on many that are) the difference in the results will be about 1%-2% at best (in most cases however it will simply be non existent). Considering the amount of money one is asked to fork out for "overclocking RAM" this result is pathetic.

Increased RAM speed does not (at all) translate to higher system performance in the way a higher clocked CPU with 33% more Cache does.

My advice for building a performance PC would be to buy 1600 CL9 RAM and the money you save compared to buying the "super-duper" 2400 RAM, that effectively gives you ah heck all, you can invest in a higher performing graphics card or CPU.

So don't be stupid and don't be fooled. The only people who are excited about faster RAM are either ignorant, salescritters or marketdroids (members of the latter two categories pretty much presupposes the first category in my professional computer experience).

And don't even get me started on the idiocy of so-called "heat-spreaders" placed over the RAM which in many cases actually INCREASES the heat of the RAM because they essentially insulate the RAM and because of the proximity to each other in the RAM slots they won't even allow for effective ventilation of the heat-spreaders themselves (hint if you cannot get your overclocked RAM to run stably try prying off the "heat-spreaders" and you could well find your problems going away).

 

[Blain]

"Performance" isn't synonymous with "better".

If I were shopping for high quality memory, I would use the following parameters as a guide...

* DDR3 rated at 1.5v or lower
* DDR3 rated at the lowest CAS I could afford
* DDR3 rated at the highest clock speed I could afford
* Limit the scope of my purchease to G.Skill, Mushkin, Samsung, Corsair XMS or Crucial (non-Ballistix)

While not wavering on the voltage point, I would balance the other issues with my budget.

Remember, my goal is not pure "benchmarking" performance, but simply finding the highest quality memory I can afford. ^_^

The only reason I pay a premium for low latency, high speed, low voltage memory is...
Quality and quality alone.

1.5v is the JEDEC DDR3 voltage standard.
Stay with 1.5v or less if you can afford it..

 

[tarmc]

it does seem like a huge waste of money getting higher than 1600 memory, only reason I have a 1866 tri channel kit is because it was dirt cheap

 

[taq8ojh]

For some reason I don't believe the OP's claims.
I specifically remember an article or some posts somewhere which explained how CL meant nothing, and raw speed was what was it all about. There were some formulas explaining how latency was directly dependant on speed, so faster module with higher CL was still better.

 

[taq8ojh]

Ha! I knew something was stinking around, and I was right.
The name was familiar, but I couldn't quite place it.
http://forums.anandtech.com/showthread.php?t=2318433

Pro tip: don't take threads from this person seriously.

 

[mikeymikec]

Didn't Anandtech do an article a while ago about unusually high RAM speeds having very little benefit on Sandy Bridge architecture?

Found it:
http://www.anandtech.com/show/4503/sandy-bridge-memory-scaling-choosing-the-best-ddr3

It makes a difference for memory benchmarks and little else, IMO.

 

[tarmc]

it doesn't really make much of a difference at all, marketing scheme. but it does seem to work, people buy the stuff.

 

[PrincessFrosty]

Well this is the memory speed, a measure of effectively bandwidth. But bandwidth with what? What actually shoves data in/out of memory?

It's the CPU, this is what reads/writes into memory so really you only need to provide enough bandwidth to make sure it's not a bottleneck for your specific CPU, above that and you won't see much or any benefit.

Faster memory is good for when you intend to overclock your CPU, or if you intend to stick with the same socket and mobo and upgrade the CPU later on, in which case it might be better investing in memory with higher bandwidth than the CPU is advertised as needing.

Timings mean very little these days, the CL timing value translates into negligible performance deltas, the only people that should care about this are people competitively overclocking and bench marking and need that tiny extra few fractions of a percent.

 

[Nec_V20]

For some reason I don't believe the OP's claims.
I specifically remember an article or some posts somewhere which explained how CL meant nothing, and raw speed was what was it all about. There were some formulas explaining how latency was directly dependant on speed, so faster module with higher CL was still better.

The person who wrote the article/post probably thought that RAM Latency was something which could be helped by Cialis or Viagra.

Latency means, "Do naff-all for the following amount of time". So at CL10 the RAM is doing nothing (albeit at at an incredibly fast rate) 12% more than RAM which has a latency of CL9.

Let's be fair about this and look at two different kits of 16GB (2*8GB) from Corsair.

In the one corner we have
Corsair 16GB (2*8GB) Vengeance Pro 1600MHz CL9 £107.98 ($168.81)
CL9 ( 9-9-9-24 )

and in the other
Corsair 16GB (2*8GB) Dominator Platinum 2400MHz CL10 £189.99 ($297.01)
CL10 ( 10-12-12-31 )

The above values in brackets refers to tCAS, tRCD, tRP and tRAS respectively and they are cumulative.

Notice the only thing which is clocked at 10 for the more expensive 2400 RAM is the CAS latency after that the latencies (i.e. do ah heck all cycles) drift even further apart.

So at the end of the day taking all the latencies into account the "spectacularly fast" 2400 RAM ends up just 18% faster than the 1600 RAM. Only a fraction of this 18% however translates directly to a higher performance of the system overall.

This is not true if the RAM can supply the data in burst mode where the 2400 RAM would be able to put its faster speed to full use over 1600 RAM. How often this is true depends on the applications which are running and manipulating RAM.

The only thing which is "spectacular" about this comparison is the difference in price which is $128.20.

Think about what you could buy which would give you a damned sight higher performance boost for the $128.20 you save by making a rational choice with regard to RAM.

Admittedly you could fart around with the latencies and voltages and MHz speed but then you would have the "heatspreaders", which actually effectively insulate the RAM, working against you. I am only going by the specification that the manufacturer will guarantee the functionality of the RAM at.

If you have money to burn then go ahead, nobody is stopping you from wasting it on "spectacularly fast" RAM. At the end of the day you will have a bit of a performance increase overall (1%-2% or none depending on which benchmark you run); however if you are on a budget the money would be better spent on other things.

 

[Nec_V20]

Well this is the memory speed, a measure of effectively bandwidth. But bandwidth with what? What actually shoves data in/out of memory?

It's the CPU, this is what reads/writes into memory so really you only need to provide enough bandwidth to make sure it's not a bottleneck for your specific CPU, above that and you won't see much or any benefit.

Faster memory is good for when you intend to overclock your CPU, or if you intend to stick with the same socket and mobo and upgrade the CPU later on, in which case it might be better investing in memory with higher bandwidth than the CPU is advertised as needing.

Timings mean very little these days, the CL timing value translates into negligible performance deltas, the only people that should care about this are people competitively overclocking and bench marking and need that tiny extra few fractions of a percent.

I would say it was the CPU and the concomitant Cache which accounts for RAM reads/writes.

However if you are building a PC on a budget then a difference of £82 would be far better spent on getting a higher performance CPU or graphics card than the 2400 RAM instead of 1600 RAM.

This is the point I am trying to make.

 

[Nec_V20]

"Performance" isn't synonymous with "better".

I 100% agree with you.

 

[sm625]

Latency is a one-off penalty. It is a unit of delay that is applied once to every operation. If you are only reading one byte at a time from RAM then yes latency matters. But when you're reading bursts of 1KB or 128KB or more, that tiny bit of latency will not matter. Since every memory read is usually done in large blocks, latency is pretty much meaningless. (Latency is huge for L1 caches, its why intel is so much faster) You can search the web and find several benchamrks where the only thing they change is the CL and it hardly affects anything.

Note the bottom two bars on this chart:

Note the higher latency actually yielded a higher max fps, while the minimums are lower. This is basically saying its all within the margin of error, ie insignificant.

click here for source

 

[glugglug]

This is why I usually shop for low latency memory. It makes more of a performance difference than the RAM clock speed, and usually only carries a small increase in price since it is less hyped. It was far more of a no-brainer back in the 90s/early 2000s, with SDR/DDR1 when we were dealing with CL2 vs. CL3 than the smaller (percentage wise) difference in latencies you can buy today, but the concept still applies.

For those saying that latency doesn't matter since "you are usually doing large block operations" you should know that the bursts reading from RAM into CPU cache are only 8 reads per burst. Even in the ideal situation that you are accessing memory in a sequential manner so the only delay between pulling in successive cache lines is CAS, it is going to apply to every 8th read. Also keep in mind that DDR3 does 4 reads per cycle once that CAS delay is finished.

So for example, for the case of DDR3 with CAS9 vs. DDR3 with CAS11, the total time to read into a cache line is CAS + 2 (8 reads in 2 cycles), which means 11 reads for CAS9 vs 13 reads for CAS11, an 18% difference. If you have random instead of sequential access so RAS and other delays come into play, the difference is going to grow even bigger.


You are never going to see such a big difference in a game, because any game engine whose working set is larger than the CPU cache is going to suck, performance wise. But there are other applications where optimizing the working set in that way is often not an option, for example compression (both general data, i.e. RAR or 7z with a large dictionary, and even more so for video compression), so you are going to see a larger impact on those applications.

 

[Nec_V20]

Ha! I knew something was stinking around, and I was right.
The name was familiar, but I couldn't quite place it.
http://forums.anandtech.com/showthread.php?t=2318433

Pro tip: don't take threads from this person seriously.

"Pro tip" really?

It is gratifying to me that my writing was of such import to you that even, after all this time, it still stuck in your memory.

If by "Pro" you mean someone who is no longer cerebrally challenged by the power switch on their own computer then I guess you would qualify.

From your replies on the other thread and here it seems that you have a monomaniacal visceral response consisting of, "I've never heard this, Nec is saying it, so Nec must be wrong". You made the mistake, and are making it now, that because I am new to this forum I must be a n00b.

You don't even try to make any kind of argument but rather resort to a well documented logical fallacy:

Ad hominem - you cannot tackle the argument so you denigrate the person making the argument.

With regard to the post you criticised in the link, it is just a fact that for lower powered CPUs what I described will dramatically improve gaming performance - end of story.

Your inglorious contribution to that thread consisted of the following and I quote:

Either troll or very stupid person identified. Identified for sure though.

Yeah and sniffing soap and drinking your own urine during full moon will cure cancer.

Shut him down someone, please. It's not amusing anymore. He seems to actually believe it.

With regard to your last statement on that thread, no I don't believe it, I know it. I know it from 18 years of experience starting with Windows NT 4. Back in the day it was known as SMP (Symmetric MultiProcessing) when it referred to more than one physical CPU on a motherboard. Now it refers to multiple cores on one CPU.

You are from the Czech Republic, so English is probably not your native language. I am from Germany and for me English is also my second language. I however make the effort to explain what I have to say as clearly as I can. Just because you cannot be bothered to read and to comprehend before you cast aspersions is a negative reflection on you rather than me.

 

[Anteaus]

I always thought the faster ram was less about performance and more about providing overhead for overclocking. I've never overclocked so take this information with a grain of salt, but I seem to remember reading about situations where the memory clock had to be increased to allow for higher stable CPU clock speeds. I have no idea if this holds true today but it seemed to make sense to me.

 

[fixbsod]

You guys don't need to come to blows over this -- it's just RAM, you want at least 1600 and with low latency. 1866 and up can give some small boosts but yeah it's not worth it due to the large price premium at the current time. Will this change in the future? Likely, but way overthinking this guys.

 

[Nec_V20]

Latency is a one-off penalty. It is a unit of delay that is applied once to every operation. If you are only reading one byte at a time from RAM then yes latency matters. But when you're reading bursts of 1KB or 128KB or more, that tiny bit of latency will not matter. Since every memory read is usually done in large blocks, latency is pretty much meaningless. (Latency is huge for L1 caches, its why intel is so much faster) You can search the web and find several benchamrks where the only thing they change is the CL and it hardly affects anything.

Note the bottom two bars on this chart:

Note the higher latency actually yielded a higher max fps, while the minimums are lower. This is basically saying its all within the margin of error, ie insignificant.

click here for source

Notice that there is a drastic drop off for the single channel RAM. There are other ways that RAM access is optimised for instance interleaving.

I took a look at all the results and in the conclusion. Just looking at the graphs you included the difference in result from

1) 1066 CAS 7 to 1600 CAS 9 was 17.83 whereas the difference between 1066 and 1600 is just over 50%

Looking at all the results there are outliers where higher MHz disproportionately registered and skewed the end result, but if one looks at all the results individually there is not so much in it that it would justify a large markup.

I would have been more impressed with the results if they had been conducted on one set of RAM or at least RAM from same manufacturer.

In some of the results the RAM which ultimately came out on top was actually behind other RAM.

 

[sm625]

You are never going to see such a big difference in a game, because any game engine whose working set is larger than the CPU cache is going to suck, performance wise. But there are other applications where optimizing the working set in that way is often not an option, for example compression (both general data, i.e. RAR or 7z with a large dictionary, and even more so for video compression), so you are going to see a larger impact on those applications.

In the cases you mention, CAS latency isnt really going to be a constraint. The precharge delay will be much more significant. Most of the overall system performance benchmarks show about a 1-3% improvement for every 20% improvement int CAS latency. A raw memory clock speed improvement of 20% will get you roughly 3-5% more performance. This is 2-5x better.


Every nongaming benchmark on this page shows speed helps more than latency

 

[Nec_V20]

In the cases you mention, CAS latency isnt really going to be a constraint. The precharge delay will be much more significant. Most of the overall system performance benchmarks show about a 1-3% improvement for every 20% improvement int CAS latency. A raw memory clock speed improvement of 20% will get you roughly 3-5% more performance. This is 2-5x better.


Every nongaming benchmark on this page shows speed helps more than latency

Personally I find this a good discussion to have because it gives people who are investing their money in building a new system a fairer overview of at least one of their choices.

What prompted me to post originally was the fact that I am annoyed at the uncritical endorsement of "faster" RAM in the media and elsewhere, the price of which is disproportionate to any real benefits.

My view is that someone on a budget should stick to 1.5v 1600 CL9 RAM and put their money into things that will give them more bang for their buck.

For instance from the graphs above getting say 16GB (2*8GB) of 1600 RAM would give far more performance benefit than buying 8GB (1*8GB) of expensive 2400 RAM.

 

[taq8ojh]

"Pro tip" really?

It is gratifying to me that my writing was of such import to you that even, after all this time, it still stuck in your memory.

If by "Pro" you mean someone who is no longer cerebrally challenged by the power switch on their own computer then I guess you would qualify.

From your replies on the other thread and here it seems that you have a monomaniacal visceral response consisting of, "I've never heard this, Nec is saying it, so Nec must be wrong". You made the mistake, and are making it now, that because I am new to this forum I must be a n00b.

You don't even try to make any kind of argument but rather resort to a well documented logical fallacy:

Ad hominem - you cannot tackle the argument so you denigrate the person making the argument.

With regard to the post you criticised in the link, it is just a fact that for lower powered CPUs what I described will dramatically improve gaming performance - end of story.

Your inglorious contribution to that thread consisted of the following and I quote:
...

With regard to your last statement on that thread, no I don't believe it, I know it. I know it from 18 years of experience starting with Windows NT 4. Back in the day it was known as SMP (Symmetric MultiProcessing) when it referred to more than one physical CPU on a motherboard. Now it refers to multiple cores on one CPU.

You are from the Czech Republic, so English is probably not your native language. I am from Germany and for me English is also my second language. I however make the effort to explain what I have to say as clearly as I can. Just because you cannot be bothered to read and to comprehend before you cast aspersions is a negative reflection on you rather than me.

You sure like to write a lot and use lots of special words to give your posts scientific-like and more serious feel. Unfortunately, it doesn't quite work.
Let me refresh your memory of what the mod who closed the thread told you:

On a side note, Nec_V20, we are (generally speaking) a "proof positive" forum. I.E. as BrightCandle puts it, "He who asserts must prove". These are after all the AnandTech forums. So go be like Anand and do some science. If you had provided rigorous and sound benchmarks alongside your commentary, this thread would have been far better for everyone involved. As it stands however you're making a somewhat wild claim (60%) without the necessary hard data to back it up.

 

[Torn Mind]

1866 Mhz RAM is faster than 1600 Mhz. It is just that the performance gains are very small and not particularly noticeable for most tests. If cost were not a variable that mattered in determining acquiring RAM, there is no reason to use slower RAM. But since there is a pocketbook penalty, slower RAM is an option because that money can be saved to purchase something else that could speed up the system more or be used on completely unrelated stuff like food, clothes, girls...

Although, an APU does benefit from faster RAM.

 

[Nec_V20]

You sure like to write a lot and use lots of special words to give your posts scientific-like and more serious feel. Unfortunately, it doesn't quite work.
Let me refresh your memory of what the mod who closed the thread told you:

As I stated in my previous reply, I am German and English is my second language.

I use the words I know, if you do not know those words or understand them then that is hardly my fault - try looking at a dictionary.

If you still don't understand then you can try this:
http://translation.babylon.com/czech/to-english/

Ignorance is not a virtue and is eminently curable.

If there is any native English speaker reading my posts and can point me to any part of any of them where the language I used was not in context or where the vocabulary I used was incorrect then I would be happy to take that under advisement.

I am not however going to apologise for my facility with regard to my second language (English) and I'm certainly not going to get a lobotomy just to fit in.

The fact that I can write a lot goes back to my beginnings 30 years ago when I got my first computer. The very first thing I taught myself was how to touch-type. I am not about to go back to hunting and pecking just because the number of words in my posts offends you.

I don't mind meeting folks half way but I draw the line at joining them at the bottom of the barrel.

You know mate there is a limit to how much you can accuse me of (and you my friend are definitely abusing the privilege) before it becomes downright offensive.

Whenever I see one of your posts the phrase, "You deprive me of solitude whilst affording no companionship whatsoever", comes to mind.

 

[Nec_V20]

1866 Mhz RAM is faster than 1600 Mhz. It is just that the performance gains are very small and not particularly noticeable for most tests. If cost were not a variable that mattered in determining acquiring RAM, there is no reason to use slower RAM. But since there is a pocketbook penalty, slower RAM is an option because that money can be saved to purchase something else that could speed up the system more or be used on completely unrelated stuff like food, clothes, girls...

Although, an APU does benefit from faster RAM.

As I said before, the difference in price between 16GB (2*8GB) 1600 CL9 RAM and 16GB (2*8GB) 2400 CL10 RAM is the difference in price between the two Haswell processors i5-4670k and i7-4770k

 

[zir_blazer]

This is why I usually shop for low latency memory. It makes more of a performance difference than the RAM clock speed, and usually only carries a small increase in price since it is less hyped. It was far more of a no-brainer back in the 90s/early 2000s, with SDR/DDR1 when we were dealing with CL2 vs. CL3 than the smaller (percentage wise) difference in latencies you can buy today, but the concept still applies.

Low Latency memory is usually also more expensive than regular JEDEC standard value memory, not sure if more or less than high Frequency. At least at the end of the DDR era when the craze for DDR 400 MHz @ 2-2-2-5 was going on, I recall that they were very, very expensive. Overally, what matters is the absolute Latency. With higher Frequency, you have worse Timmings, but the overall Latency usually goes down, not up, plus you get the Bandwidth. This means that on average, even with higher Timmings, you end up having more memory performance anyways.

AMD APUs's GPUs DO like high Frequency RAM, they make use of all the Bandwidth that they can get. Trinity scaled well even up to Dual Channel 2400 MHz. I suppose that as Intel adds bigger integrated GPUs it will also apply for them (Don't recall if for Haswell it already does). For CPU only, it seems that with Dual Channel 1333/1600 MHz you're pretty much maxed, and either more Bandwidth or lower Latency barely affects performance.


I don't like the RAM shopping habits of the people here (I'm looking to Blain, that on every Post copypastes the same three Quotes!). When I shopped for my current RAM, the ONLY thing I cared about was the ICs that they used, not the default specs of the RAM. That's how I ended up with some AMD branded memory that uses Hynix MFR ICs, that are know to be among the top choice in current high end modules, and best of all, they also were among the cheapest 2 * 8 GB 1866 MHz kits.
If you know that on average an IC does good or scales in some way that you think you can use, you could have realist expectatives about how you want it to run in your machine, in the same way that people always ask about average overclocks for their new Processor so they can easily get there. I think that knowing that gives you the best possible choices, but is actually much harder to get it right.


Also, according to Newegg, the cheapest 2 * 8 GB 1600 MHz kit is 109 U$D, but its from an unknow brand, know ones start at around 120 U$D. For 1866 MHz, they start at 120 U$D, and you can already get know ones at the 130-140 U$D range. Actually, this GSkill Sniper 1866 MHz CAS 9 kit cost 117 U$D with the 15% discount. Its even cheaper than popular 1600 MHz kits. At the time, I even almost purchased that one, but I didn't like the fact that it appeared to use some Nanya ICs that no one knew what they were capable of, reason why I picked the AMD ones.

 

[Ketchup]

Interesting discussion. Nec_V20. I feel this discussion is 13 years too late. Why? In 2000, SDRAM was in virtually every computer you could buy, and the last time I ever remember cas latency and speed causing a noticeable difference in my computer's performance (noticeable to me, not just in benchmarks).

Heck, the laptop I am on right now, Lenovo Thinkpad W530 with 8GB DDR3, is running one stick of RAM, and is one of the faster machines I have ever used. If it had a faster hard drive, like in the desktop in my sig or an SSD, it would be the fastest (and my desktop does run dual channel).

I think you need to come to grips with the times here. And, as others have stated, making statements as you have done would require some proof, which you don't seem to have.