DDR3 Performance: What Makes Memory Perform Better?

[JackNaylorPE]

When DDR3-1600 first hit the shelves, just about all of it was 1.65 volts. As production lines and yields improved, improvements allowed that voltage to be reduced to what we see now. That being said, I have never seen a documented incident where someone fried anything at 1.65 volts.

And for those that still believe that your warranty is voided at > 1.5 volts, take a look at Intel's XMP page.

http://www.intel.com/content/www/us/en/gaming/extreme-memory-profile-xmp.html

XMP is overclocking and just as with CPUs and GPUs higher speeds than "stock" require more voltage. When you buy RAM that says 2400 on the package, that is the "overclocked" speed and it will require higher than 1.50v.

Intel has **no issue with this whatsoever** as is clearly indicated if you click on the above link and look at the RAM sets that have been certified as compatible with Intel i5 / i7 CPUs. About half of the sets on that list are > 1.5 volts.

I'd love to see Intel argue that you somehow voided your warranty when you are using a set of RAM on their compatibility list when the list states the set was tested and certified compatible by Intel for the i5 / i7 at 2400 at 1.65 volts.

How high you can go will depend mostly on the manufacturer of the modules used ..... Hynix modules are the most tolerant .... search around and you will find overclockers running 1.9 volts and more.....even higher on DICE and LN2

Mixing sets can bring mixed results but like anything else performance and compatibility depends on how "like" the modules are. This has nothing to do with the brand name ... which basically defines the shape and color of the heat sinks more than anything else..... but by what modules are used underneath.

For example.... let's say
you have 2 x 8GB Corsair Vengeance Pro (Version 4.40)

Now 18 months later you wanna take a shot at going to 32 GB... you find 2 sets available at a decent price ....which set do you get ?

a) 2 x 8GB Corsair Vengeance Pro 2400 CAS 10 (Version 4.72)
b) 2 x 8GBMushkin Redline 2400 CAS 10 (Any Version)

The answer is b) and the reason is that Corsair switched memory module suppliers after version 4.51 The Corsair VPs (pre version 4.51) and Mushkin RL's both use the same Hynix modules and having identical same spec modules will mean more than whose name / logo is painted on the heat sink.

The general difficulty with RAM comparisons, and this is for the reader rather than the author, is that it is typically a "this is what they offered" kinda thing which results in a mixed bag of performance / quality levels which, despite and effort by the author to dissuade such, tempts the reader to draw conclusions based upon brands rather than model line.

As for gaming, the effect of RAM Speed can be small or large depending on the game as THG showed whereby the difference was 0% in Metro 2033 and 11% in F1 going from 1600 to 2400. But like anything else, performance is constrained by the weakest link and if you want to see how RAM impacts gaming performance .... you really need to use multiple GFX cards, overclocked CPUs and focus more on min. fps than average fps.

 

[Tradesman1]

Mixing sets can bring mixed results but like anything else performance and compatibility depends on how "like" the modules are. This has nothing to do with the brand name ... which basically defines the shape and color of the heat sinks more than anything else..... but by what modules are used underneath.

For example.... let's say
you have 2 x 8GB Corsair Vengeance Pro (Version 4.40)

Now 18 months later you wanna take a shot at going to 32 GB... you find 2 sets available at a decent price ....which set do you get ?

a) 2 x 8GB Corsair Vengeance Pro 2400 CAS 10 (Version 4.72)
b) 2 x 8GBMushkin Redline 2400 CAS 10 (Any Version)

The answer is b) and the reason is that Corsair switched memory module suppliers after version 4.51 The Corsair VPs (pre version 4.51) and Mushkin RL's both use the same Hynix modules and having identical same spec modules will mean more than whose name / logo is painted on the heat sink.

Ideally this could be true, but in real life, not so much. This is addressed more fully in the upcoming followup article, but simply: while true you can get different DIMMs from different manufacturers to play at times, it's not as simple as what you portray. If by modules you refer simply to the memory chips, when manufactured those chips are tested and often broken up (binned) to different specs, so while you may see memory chip XX used in different DIMMs you have no idea of either the chip's manufacturer's binning nor the DIMM manufacturers binning of said chips. Then too there are the DIMMs themselves, the PCB used, the solder, the SPD programming and other play also as different DIMM manufacturers don't all use the same exact products, and even if they did, those same products when from different production lots can affect things.

This was touched on in the a article where two sets of 2x8GB sticks from the same manufacturer wouldn't play right out of the box. This is why it's always best to get the full amount of DRAM you want in a single package which ensures the sticks have been tested to all play nice. You could pull consecutive sticks right off the assembly line that won't play.

 

[ddpruitt]

The latency numbers presented are utterly useless. Typically you can tighten the timings when running RAM at a lower speed and in most cases they run at the same latency in terms of time. So unless we know what the timings were set to and why it's not helpful. It also makes me leery of the other results, a larger latency will also affect the other numbers.

You also have to be careful with the multitasking numbers. I've seen even tiny changes in local timings double global effects.

The differences are interesting, I'd like to see how these affect something like video rendering and encryption.

 

[Tradesman1]

The latency numbers presented are utterly useless. Typically you can tighten the timings when running RAM at a lower speed and in most cases they run at the same latency in terms of time. So unless we know what the timings were set to and why it's not helpful. It also makes me leery of the other results, a larger latency will also affect the other numbers.

You also have to be careful with the multitasking numbers. I've seen even tiny changes in local timings double global effects.

The differences are interesting, I'd like to see how these affect something like video rendering and encryption.

___________________________

Good points and could be the source for an additional followup, this was written to explore the differences in how DRAM runs between the two platforms in a typical 'normal' user scenario. I chose 2400 as I knew the AMD rig could run a full load at 2400 (was running the Tridents when I did the build. Since then I've continued to experiment and have been able to take them a bit higher with mixed results, same on the Z87 as well as my Z97, and again mixed results (varying data rates at varying timings). Same thing w/ downclocking the data rate and tightening the timings (but that hasn't really shown anything in the way of performance gains). I've also played with the sets at 2400 as far as the primary and secondary timings and can pull performance gains there with some sets. Unfortunately, adding these type of things would have resulted in an overly cumbersome article, and (as a guesstimate) tripled or more the time involved, and that might be a conservative estimate as I would then want to extend that into 1, 2, 3 stick setups as well as how these setups affect the cpu OCing......and well, think you get the idea, one thing leads to another.

I appreciate the comments and will try to look into what I/We can do, I'd guess something maybe testing on a single platform (or separate articles for each) and possibly with a smaller sampling of sets. Some have already been returned.

 

[Crashman]

Hey guys, remember that the 1.65V limit popped up in reference to a specific processor after a die shrink and a less-stable doping material. I think the difference had to be a maximum of 0.50V between the memory controller voltage and DRAM voltage. Was it early Ivy Bridge cores that had the problem? It's all water under the Sandy Bridge now

 

[lordmogul]

Like I wrote on the UK release of the same test:

I would also like to see how single vs dual channel performs nowadays in real time applications.
Often in those RAM tests they run 8 GB on single channel mode (based on the 8 GB modules used)
And mostly I see only running synthetic benchmarks like aida and sandra, what about the "daily" usage? (video rendering, heavily tabbed webbrowsing, gaming, video and photo editing, office, file transfers, watching HD/UHD videos?

 

[Vogner16]

I really liked this read. bottom line as I have been telling people for years is speed over size. the 16 gb of 2400 out performed the more expensive 32 gb of 1600 mhz... wow. and how often do you actually use 32 gb of ram! im not a video editor and I don't own a game that even uses the 16 gb I have. great job with this article!

 

[rolli59]

I really liked this read. bottom line as I have been telling people for years is speed over size. the 16 gb of 2400 out performed the more expensive 32 gb of 1600 mhz... wow. and how often do you actually use 32 gb of ram! im not a video editor and I don't own a game that even uses the 16 gb I have. great job with this article!

You are right up to the point when applications run out of ram and have to start using the swap/page file on the HDD to supplement ram. Therefore quantity matters more than speed in overall use of the computer!

 

[Tradesman1]

I really liked this read. bottom line as I have been telling people for years is speed over size. the 16 gb of 2400 out performed the more expensive 32 gb of 1600 mhz... wow. and how often do you actually use 32 gb of ram! im not a video editor and I don't own a game that even uses the 16 gb I have. great job with this article!

You are right up to the point when applications run out of ram and have to start using the swap/page file on the HDD to supplement ram. Therefore quantity matters more than speed in overall use of the computer!

_______________________

There's valid arguments displayed on both sides here, speed vs quantity, which plays in with previous posts of things people would like to see Have been toying w/ ideas for testing methodology to where we might come up with some quantifiable results...stay tuned....and suggestions are welcome

 

[rolli59]

My point is, you will have to have enough ram to run the applications you use, to little of faster ram hurts more than the speed penalty between enough of say 1600 vs 1866.

 

[andre_888]

What I would like to see, or maybe I should go ahead and test myself, are the following relationships in memory benchmarking :

NB frequency
Timings
and DRAM frequency.

 

[Vogner16]

I really liked this read. bottom line as I have been telling people for years is speed over size. the 16 gb of 2400 out performed the more expensive 32 gb of 1600 mhz... wow. and how often do you actually use 32 gb of ram! im not a video editor and I don't own a game that even uses the 16 gb I have. great job with this article!

You are right up to the point when applications run out of ram and have to start using the swap/page file on the HDD to supplement ram. Therefore quantity matters more than speed in overall use of the computer!

while running out of ram is a situation where performance hits will be extreme, only a very select few programs will ever hit these bounds as even a computer with 4 gb of ram will not see this hit while playing most games. I always put 8 gb of ram in computers and the fastest speed possible. never have I ever seen a gaming computer take a performance hit due to maxing out the ram on the system when 8 gb are installed. and on a side note I haven't even seen this with 4 gb computers... so while size of ram is needed per say if you had 2 gb in your computer you might max it out when doing regular day tasks, but with 4 or 8 gb which is standard in most every computer out there you will not see this hit.

for these reasons speed will give you more performance than size for the dollar spent.
_______________________

There's valid arguments displayed on both sides here, speed vs quantity, which plays in with previous posts of things people would like to see Have been toying w/ ideas for testing methodology to where we might come up with some quantifiable results...stay tuned....and suggestions are welcome

 

[JackNaylorPE]

Ideally this could be true, but in real life, not so much. This is addressed more fully in the upcoming followup article, but simply: while true you can get different DIMMs from different manufacturers to play at times, it's not as simple as what you portray. If by modules you refer simply to the memory chips, when manufactured those chips are tested and often broken up (binned) to different specs, so while you may see memory chip XX used in different DIMMs you have no idea of either the chip's manufacturer's binning nor the DIMM manufacturers binning of said chips. Then too there are the DIMMs themselves, the PCB used, the solder, the SPD programming and other play also as different DIMM manufacturers don't all use the same exact products, and even if they did, those same products when from different production lots can affect things.

This was touched on in the a article where two sets of 2x8GB sticks from the same manufacturer wouldn't play right out of the box. This is why it's always best to get the full amount of DRAM you want in a single package which ensures the sticks have been tested to all play nice. You could pull consecutive sticks right off the assembly line that won't play.

Yes, it's not guaranteed ... you said that in the article, I said it in my comments so I don't understand where the disagreement is.

And I am speaking from "real life" experience .... I have had success > 80% of the time.... I posted previously on the instance where I bought 2 sets of GSkill to upgrade 2 different user machines. One had a pair of Mushkins, one a pair of Corsairs. Nothing I could do (within reasonable limits) made the Gskill work with either the Mushkin or Corsairs.... but the Corsair and Mushkin worked fine together and the two GSkills worked together on 1st boot w/ no adjustments

In this instance was speaking to a set of say Corsair 2400 w/ 10-12-12-28 timings (indicative of Hynix modules) being mixed with a set of Mushkin 2400 w/ 10-12-12-28 timings (again indicative of Hynix modules) as opposed to the new Corsair 2400 w/ 10-12-12-31 timings (indicative of other then Hynix modules). I never said it was 100% guaranteed ... Of course it's a crapshoot which is why I used the term " take a shot". "Making an attempt" can in no logical way be interpreted as suggesting guaranteed success.

Now 18 months later you wanna take a shot at going to 32 GB

http://dictionary.reference.com/browse/take+a+shot+at

35. have / take a shot at, make an attempt at:

In the past 25 years of building PCs, I always have recommending "if ya can't decide between X GB and 2X GB, play it safe and go large", but sometimes users make the "budget decision" and then change their mind. If it becomes a choice of continuing to use a sluggish PC and "taking a shot" at solving some of the problems, many folks will make the attempt.

But binning is exactly the reason why this is true. When a set of 4 sticks is tested and runs at 1.65 volts, it goes in the package. If it doesn't, say it tests stable at 1.67 volts, it can't go on the package cause it it needs 1.67 volts and the package says 1.65. Broken into sets of 2, those same sticks may both pass and be sold at that speed and timings. So what would happen if you bought those two packages and ran them at 1.67 v .... they'd work, they just didn't meet the spec'd 1.65 limit. If they are really far off, one of the sets might get labeled and sold a lower speed or timings.

It's also important to note whether the RAM pairings / testing are done "on the line" or actually hand tested. Only ones that I know do hand testing are Mushkin and Avexir. Not saying no one else does just that I never saw it documented. I always assumed that Corsair Dominators were given their price premium but again, never saw it documented. (If anyone has a reference in this regard, would appreciate getting educated)

The more things you have the same, the better your chance of success. Same with PSUs..... The Corsair HX 650 unit made by Seasonic will bear more similarities to the Seasonic unit made on the same platform that it will to the Corsair CX750. Same with GPUs .... Two different brand 780s with Samsung RAM, are more similar from the memory usage / overclocking perspective than two Asus 780s with one being Eplida and one being Samsung.

Ya matched rated speed, that's a start ....
Ya matched rated timings, I'm starting to feel lucky ...
Ya matched modules supplier and part number, I'll bet on that ...

Corsair's VPs w/ Hynix are a more difficult match than Corsair VPs with different timings and a less expensive brand of memory module. Two sets, different brand w/ same native speed, same native timings, and same module part number, while by no means guaranteed, is just a better bet. When I was buying my Asus 780s, the one thing I feared more than getting two with Elpida memory was getting one with Elpida memory.

 

[logainofhades]

I have always run my 2133 ram, at 1866. I am not good at tweaking memory, as I have never taken the time to learn the ins an outs of it. It never seemed to want to run at 2133, for some reason. I would get lockups, or the system would reboot. Set it to 1866, and runs just fine. I have this set. http://www.newegg.com/Product/Product.aspx?Item=N82E16820231653

 

[Evolution2001]

One area I'm still curious about, and I didn't see it referenced in any of the 60+ comments nor the article itself... "Heat and Heatsinks".
Long ago when DDR3 started shipping, I had heard that because it runs at lower voltages, the whole heatsink thing was really a non-factor, as the chips simply don't run as hot as previous gen chips.
Is this true now? Was it ever true? You can obviously buy 'value' RAM from most of these vendors, and those sticks don't have heatsinks.
Is there an actual benefit to having heatsinks? Would the chips get hot enough on their own to fail stress testing?

 

[InvalidError]

Long ago when DDR3 started shipping, I had heard that because it runs at lower voltages, the whole heatsink thing was really a non-factor, as the chips simply don't run as hot as previous gen chips.
Is this true now? Was it ever true?

No baseline DRAM spec has ever required heatsinks. Standard DRAM chips use less than 0.5W a piece and that can be taken care of by the PCB's internal copper planes.

Desktop DRAM only needs heatsinks when you start running it at non-official speeds and increased voltages. If you look at low-profile DIMMs, they have effectively no heatsink to speak of since their heat-spreader provides almost no extra surface area compared to the DIMMs themselves. They may even make things slightly worse by reducing clearance (and airflow) between DIMMs.

 

[Tradesman1]

JackNaylorPE, sorry your response was to long winded to copy the whole thing, but it's basically explained in part of my response you quoted, that being

"This was touched on in the a article where two sets of 2x8GB sticks from the same manufacturer wouldn't play right out of the box. This is why it's always best to get the full amount of DRAM you want in a single package which ensures the sticks have been tested to all play nice. You could pull consecutive sticks right off the assembly line that won't play."

I'm thrilled you've had this 80% success rate, but the fact remains, when you can take identical model sets from the same manufacturer and they often will not play (not to mention the manufacturers won't guarantee that mixing sets will play nice), then I can't (and I don't know of any reputable builders or IT folks (including manufacturers of DRAM, mobos, even CPUs)) that will advocate going out and buying additional DRAM with the express purpose of mixing it with other DRAM. If you already have additional sticks, then yes, go ahead and try it. (This is something I do all the time with sticks on hand). Otherwise give thought to if they don't play, you may be looking at RMA, paying returned shipping, restocking fees etc. Even with the same exact models we see this all the time where the 'smart' shopper buys two packages of 2x4 or 2x8GB rather than buying the 4 stick guaranteed package - to save $5-10 , only to find the two packages don't play

 

[InvalidError]

Even with the same exact models we see this all the time where the 'smart' shopper buys two packages of 2x4 or 2x8GB rather than buying the 4 stick guaranteed package - to save $5-10 , only to find the two packages don't play

Even if you buy sets of quads, those sets are only "certified" on a limited number of motherboards and CPUs, so it is still possible to run into cases where it might not work even under what would be hypothetically best-case circumstances.

Signal integrity on large multi-drop busses carrying GHz+ signals is tricky and it only gets worse when you add board-to-board connectors, which is why all other interfaces have moved to point-to-point high-speed serial. DRAM is still using parallel due to the difficulty of integrating high-speed logic on fabrication processes optimized for memory chips.

 

[Tradesman1]

Even with the same exact models we see this all the time where the 'smart' shopper buys two packages of 2x4 or 2x8GB rather than buying the 4 stick guaranteed package - to save $5-10 , only to find the two packages don't play

Even if you buy sets of quads, those sets are only "certified" on a limited number of motherboards and CPUs, so it is still possible to run into cases where it might not work even under what would be hypothetically best-case circumstances.

Signal integrity on large multi-drop busses carrying GHz+ signals is tricky and it only gets worse when you add board-to-board connectors, which is why all other interfaces have moved to point-to-point high-speed serial. DRAM is still using parallel due to the difficulty of integrating high-speed logic on fabrication processes optimized for memory chips.

______________

The example I gave is that 4sticks in a given package are guaranteed to play together where as if you buy two sets of two sticks of the same model - there is no guarantee they will play nice together.

To my knowledge DRAM isn't 'certified' to a given CPU, the memory manufacturers and mobo manufacturers each establish compatibility list of mobo to DRAM, but it's usually limited to either the mobos or the DRAM they have available to test with (as well as time to test) and often put a disclaimer out to the effect of, it's a limited tested sample, others are also compatible

 

[InvalidError]

To my knowledge DRAM isn't 'certified' to a given CPU, the memory manufacturers and mobo manufacturers each establish compatibility list of mobo to DRAM, but it's usually limited to either the mobos or the DRAM they have available to test with (as well as time to test) and often put a disclaimer out to the effect of, it's a limited tested sample, others are also compatible

And that was the point I was trying to make: the "certification" on quad kits has extremely narrow applicability compared to all the possible CPU, motherboard and DRAM configurations out there.

Since the IO drivers and receivers connected to the DRAM bus are in the CPU, they are subject to all the usual CPU yield variations, variations between die models, die revisions, product generations for a given socket, etc. on top of motherboard and DIMM production variations.

 

[GObonzo]

never had any trouble running my 2400 G.Skill Ripjaws Z on AM3+, so don't be discouraged. worked great with my 8350 @ 4.8GHz.

 

[Tradesman1]

To my knowledge DRAM isn't 'certified' to a given CPU, the memory manufacturers and mobo manufacturers each establish compatibility list of mobo to DRAM, but it's usually limited to either the mobos or the DRAM they have available to test with (as well as time to test) and often put a disclaimer out to the effect of, it's a limited tested sample, others are also compatible

And that was the point I was trying to make: the "certification" on quad kits has extremely narrow applicability compared to all the possible CPU, motherboard and DRAM configurations out there.

Since the IO drivers and receivers connected to the DRAM bus are in the CPU, they are subject to all the usual CPU yield variations, variations between die models, die revisions, product generations for a given socket, etc. on top of motherboard and DIMM production variations.

_________________

Where the mobos and DRAM are determined to be compatible with each other is simply that - they will work together. As was pointed out in the article not all CPUs can handle all data rates of DRAM available. While Yes, say the ACME 3000 DRAM will run on the UNKNOWN Z97 - you still need a CPU that can carry the sticks at 3000. This is a point that is also addressed in the followup article - all too often you see a person say "well, the mother board can run 2800 so just get that", (which happens all the time), but they seldom mention, on the fact side "if however you want to run those expensive 2800 sticks at 2800 get a CPU that can run them, else you may have an entry level CPU that can't run them at more than 1600". This comes from a mindset dating back a few years where the MC (memory controller) was in the mobo chipset.

The whole point here though is that buying a 4 stick package you have a guarantee that the four sticks will play together, buying two packages of two sticks of the same model of DRAM there is no guarantee that all four of those sticks will play.together.

 

[ah]

I'd like you to test the DDR3-2400 32GB vs the DDR3-2400 8GB with an SSD card installed to see which one will come out on top. Thank you.

 

[Tradesman1]

Both rigs used a 256GB SATA iii Phoenix Pro SSD.