DDR3 Performance: What Makes Memory Perform Better?

[Bannereus]

I think the multitasking simulation is a great idea for a benchmark, more relevant to an average user than an absolute bandwidth or latency

Question for your next one:
How significant is not just the capacity of the memory, but number of DIMMs, and motherboard 'channels'?

ex Does 16GB DDR3-1600 *noticably* faster/slower in 1x16GB as 2x8, or 4x4? Or 4x4GB in a dual-channel vs triple or quad channel?
Is there a *significant* power draw difference between 1x16 and 4x4, that may strain a lesser motherboard (or affect power supply choice)?


(Planning a platform upgrade after Skylake hits the desktop, still don't know as much about RAM as I'd like to)

 

[Tradesman1]

I think the multitasking simulation is a great idea for a benchmark, more relevant to an average user than an absolute bandwidth or latency

Question for your next one:
How significant is not just the capacity of the memory, but number of DIMMs, and motherboard 'channels'?

ex Does 16GB DDR3-1600 *noticably* faster/slower in 1x16GB as 2x8, or 4x4? Or 4x4GB in a dual-channel vs triple or quad channel?
Is there a *significant* power draw difference between 1x16 and 4x4, that may strain a lesser motherboard (or affect power supply choice)?


(Planning a platform upgrade after Skylake hits the desktop, still don't know as much about RAM as I'd like to)

___________

Thanx for the comments, and yes I will be addressing both the issues raised

 

[Victor Bettenhausen]

I've been running 24gb (6x4gb) Mushkin 1600 in my system for three years at 1.7v. Which incidentally, is the recommended voltage according to Mushkin. Intel is very conservative, historically. As long as you don't get ridiculous with your voltage settings you're not likely to cause any damage to your cpu, or memory

 

[Tradesman1]

I've been running 24gb (6x4gb) Mushkin 1600 in my system for three years at 1.7v. Which incidentally, is the recommended voltage according to Mushkin. Intel is very conservative, historically. As long as you don't get ridiculous with your voltage settings you're not likely to cause any damage to your cpu, or memory

______________

Yes, I like to keep to spec, at time a little more is needed, 1.7 is about my upper limit, other than Extreme OC or Suicide Shots. I do however stay away from lower data rate DRAM that calls for higher voltage, like 1600 sticks at 1.6, 1.65, generally it can be indicative of using weak memory chips. About 95+ % of all 1600 sticks take 1.5 or less, even high performance

 

[Karsten75]

Nice article to use when I want to link a question to the effect of RAM in a system - both from a perspective of the quantity and speed thereof.

 

[Tradesman1]

Nice article to use when I want to link a question to the effect of RAM in a system - both from a perspective of the quantity and speed thereof.

______________

Thanx, at present for the followup article, I thinking I will have about 30 points I'll go into, plus whatever suggestions come up here. It's tentatively titled "DDR3 - FAQs and Fiction". Am hoping they will be able to publish with each point/item having a link so those in the forum can link to the individual answers/explanations to them, a few examples are :

Will a quad channel set of DRAM run in my dual channel motherboard?
Your motherboard supports 3200 DRAM so you can use any DRAM you want.
It’s DDR3, all DDR3 is the same.
Which is better – low CL or high data rates?
8GB is all anyone needs (or 8GB is all you’ll need for the next X Years).

Have a diverse selection now, but always looking for more questions 😉

 

[shrapnel_indie]

Bummer most of the faster memory seems to be EOL stock. Leaves me wondering if there will be newer models from vendors soon or if we are out of luck and stuck with whatever we can get. Case in point. I've been looking at the Team 16GB kits with a CL11 rating at 2666... All gone now... most show no longer available at Newegg. the first was the black modules, and now more recently the white modules (although still not showing no longer available, just out of stock.) I guess for my price-point I may be forced down to 2400 and CL10. Ugh! Not fun when you want to have a new build but have to piece-meal the parts for a specific platform.

According to the article here, I guess the best value in this line of "tests/benchmarks" (using the term loosely as it was pointed out it wasn't a normal review) is the Team DDR3-2400 CL10 kit (link points to the 16GB kit btw, it would be a steal for that price if it was a 32GB kit.)

 

[Tradesman1]

Bummer most of the faster memory seems to be EOL stock. Leaves me wondering if there will be newer models from vendors soon or if we are out of luck and stuck with whatever we can get. Case in point. I've been looking at the Team 16GB kits with a CL11 rating at 2666... All gone now... most show no longer available at Newegg. the first was the black modules, and now more recently the white modules (although still not showing no longer available, just out of stock.) I guess for my price-point I may be forced down to 2400 and CL10. Ugh! Not fun when you want to have a new build but have to piece-meal the parts for a specific platform.

According to the article here, I guess the best value in this line of "tests/benchmarks" (using the term loosely as it was pointed out it wasn't a normal review) is the Team DDR3-2400 CL10 kit (link points to the 16GB kit btw, it would be a steal for that price if it was a 32GB kit.)

______________

I saw that, in my article/review I did a recap of each set of DRAM individually, the staff changed it to the advertising type thing shown, the ADATA sticks should also indicate a 'X2' as they sent 2 - 2x8GB sets, they also link the Corsair to a 2x8GB set when they actually sent me a full 4x8GB set, for those at NewEgg they may simply be out of stock...but not sure. Believe that's the case with the Team (white) Xtreem.

Either the Team or the Tridents in 2400/10 will normally run at 2666/11 with a CPU OC and a bit of extra voltage to the DRAM or the MC (memory controller.

 

[damric]

I've been known to abuse more than 1.7v for DDR3. I think that's still within margin of error.

Hell I've had some old OCZ Reapers DDR3 that were 1.85 stock.

 

[Tradesman1]

and the GSkill Flares (think it was F3-16000CL7D-4GBFLS, circa 2010/11) , DDR3 2000 sticks at CL7 - took those to 2400/8 at 1.7, those were some really sweet sticks!

 

[f-14]

Could you revisit the SSD versus more RAM testing, where there's less wear on the SSD with more memory?

there's no need it was already proven more ram equals less accessing on the SSD since you have a larger swap file via increased ram 64>32>16>8>4>2>1 hands down every time.

pretty much confirms everything i thought about ram timings and performance, more memory even at mid speed always tops less memory at top speeds.

i'm trying to think of a game that would be replicable for all configurations that would saturate the ram. all i can think of is games like starcraft, or total annhilation and falcon4 from back in the edo/sdram days when you recorded the games. nearest i can think today would be starcraft2 and supreme commander 1 or 2 and DSC: A-10 Warthog A or C
_____________

True, I was considering this though in conjunction with an article on optimization of SSDs. The Swap/page file is generally located on a hard drive, so not only will you have less reads and writes to the page/swap file by having more DRAM - you could place the swap/page file on a RAMDisk as an option, just a possible 😉

 

[Sabishii Hito]

Bummer most of the faster memory seems to be EOL stock. Leaves me wondering if there will be newer models from vendors soon or if we are out of luck and stuck with whatever we can get. Case in point. I've been looking at the Team 16GB kits with a CL11 rating at 2666... All gone now... most show no longer available at Newegg. the first was the black modules, and now more recently the white modules (although still not showing no longer available, just out of stock.) I guess for my price-point I may be forced down to 2400 and CL10. Ugh! Not fun when you want to have a new build but have to piece-meal the parts for a specific platform.

According to the article here, I guess the best value in this line of "tests/benchmarks" (using the term loosely as it was pointed out it wasn't a normal review) is the Team DDR3-2400 CL10 kit (link points to the 16GB kit btw, it would be a steal for that price if it was a 32GB kit.)

______________

I saw that, in my article/review I did a recap of each set of DRAM individually, the staff changed it to the advertising type thing shown, the ADATA sticks should also indicate a 'X2' as they sent 2 - 2x8GB sets, they also link the Corsair to a 2x8GB set when they actually sent me a full 4x8GB set, for those at NewEgg they may simply be out of stock...but not sure. Believe that's the case with the Team (white) Xtreem.

Either the Team or the Tridents in 2400/10 will normally run at 2666/11 with a CPU OC and a bit of extra voltage to the DRAM or the MC (memory controller.

The main issue is that SK Hynix apparently no longer manufactures the 4Gbit MFR ICs that were used for high density kits (2x8GB, 4x8GB) rated 2666C11 or occasionally 2400C10 (which is actually a tougher spec to meet). I don't think G.Skill ever used them for the Trident X 2400C10 modules but opted for Samsung 4Gbit chips which can't do much more than 2550.

 

[rolli59]

Nice article, looking forward to the follow up but would like to see if we can get numbers on any effect from increased ram speed on gaming performance as well.
_____________
Good point, I looked at this for inclusion, but was more centered on a general AMD vs Scenario here as far as overall performance. In gaming there are a number of variables that play in. In gaming DRAM as long been not much more than a data pipeline, so increasing the data rate has resulted in minimal gains as far as FPS and gaming enjoyment. Today though developers are more and more utilizing DRAM for performance increases, but it's (at this point) specific game oriented. To simplify, more are opening up the ability to use more DRAM, which allows more data to be maintained in DRAM rather than written to a page/swap file, so in effect faster data rates and the amount of DRAM are playing in more. Think this would be a topic for a separate article and examine some specific games and upcoming trends in game development

 

[chechak]

frequency has such a measurable impact on compression/other but in gaming does it effect to much of FPS ?
typical gamer /pro gamer

 

[shrapnel_indie]

In gaming there are a number of variables that play in. In gaming DRAM as long been not much more than a data pipeline, so increasing the data rate has resulted in minimal gains as far as FPS and gaming enjoyment. Today though developers are more and more utilizing DRAM for performance increases, but it's (at this point) specific game oriented. To simplify, more are opening up the ability to use more DRAM, which allows more data to be maintained in DRAM rather than written to a page/swap file, so in effect faster data rates and the amount of DRAM are playing in more. Think this would be a topic for a separate article and examine some specific games and upcoming trends in game development

That would be interesting. ESPECIALLY to those of us who dabble in game design (or are looking to go pro/indie.)

 

[Tradesman1]

Rolli59, Chechak & Shrapnel_Indie

Will check what's on tap and with the powers that be and see if we can do an article on this 😉

 

[Banned4noreason]

I'm curious why hardware review sites never report (or run) multiple tests or run statistical comparisons. With results so close as these, surely many are within the margin of error for the measurement and testbed.

 

[Crashman]

I'm curious why hardware review sites never report (or run) multiple tests or run statistical comparisons. With results so close as these, surely many are within the margin of error for the measurement and testbed.

We do averages when it matters, but the problem is those darned 1-second differences, where 1000 tests might round it to 33 seconds (say, 33.4449 but the 1001st test rounds it to 34 (to 33.5000).

 

[Crashman]

Rolli59, Chechak & Shrapnel_Indie

Will check what's on tap and with the powers that be and see if we can do an article on this 😉

One of the "powers that be" would love to see the type of article that can find an award-worthy distinction...I'm speaking of the power who writes the "Best DRAM" awards showcase

 

[InvalidError]

We do averages when it matters, but the problem is those darned 1-second differences, where 1000 tests might round it to 33 seconds (say, 33.4449 but the 1001st test rounds it to 34 (to 33.5000).

If people are making a big deal out of a single second difference, then you can simply add a decimal place or two to better represent how small of a difference there really is.

But when averages are that close or are at a level where they should not matter anymore, it becomes more interesting to look at how well the average represents the data set. That's what you have standard deviations or Nth percentile for. If it is possible to accurately line up frame-time data between benchmark runs, then it would be possible to add frame time variance analysis on a per-point basis to show how reproducible results are on a given platform, GPU and game settings.

 

[Tradesman1]

I'm curious why hardware review sites never report (or run) multiple tests or run statistical comparisons. With results so close as these, surely many are within the margin of error for the measurement and testbed.

------------------------------------------------------------

If I understand what you are looking for, each of the results came as a result of running the same tests 5 times, each from a clean boot, on each set of DRAM in the config indicated, the high and low scores were thrown out from possible anomalies and the remaining three scores were averaged to become the 'result' for that set and test.

 

[MeteorsRaining]

This is one of the best DRAM related articles I've read. Great to see the issues of incompatibility by mixing sets are being emphasized on. The "simulation" test is a good measure for quantifying real world differences of different DRAM speeds and capacity.

A lot of work hardwork has been put in this, and it's totally worth for such an eye-opening article that this is. I fancy linking this to solve many related issues which are raised in the forums on a daily basis. Looking forward to more articles like this! Great work!

 

[Banned4noreason]

We do averages when it matters, but the problem is those darned 1-second differences, where 1000 tests might round it to 33 seconds (say, 33.4449 but the 1001st test rounds it to 34 (to 33.5000).

But when averages are that close or are at a level where they should not matter anymore, it becomes more interesting to look at how well the average represents the data set. That's what you have standard deviations or Nth percentile for.

Yes, this is what I'm getting at. When testing things like DRAM and motherboard timings or the same GPU by different vendors, results are so close. Often, however, one of them is award the best performer, or proclaimed the fastest. But without displaying any representation of sample variation (either for the sample piece of equipment on multiple tests OR different physical "samples" of the same item) it is hard to know if, truly, one IS faster than the other.
________________

I see what you are saying, and yes it would be nice (could just about write an article in reponse to this question 😉 ). This is something I enjoyed in doing this piece as I was 'testing' on two completely different platforms (I also ran many of the tests on my Z97. Among the things to be considered is getting the samples for testing, I was surprised all the manufacturers but one responded w/ samples at all, I think if I had asked for multiple sets of DRAM of the same model, the response would have been smaller. There's often a time constraint on how long to get an article written which limits the time that testing can be done. If testing multiple samples against another brand, if one dies or is flawed, do you continue and test in say a three items of A against 4 items of B. Then too, if as Crashman indicated a sample or two throws off the entire end result, how do you quantify results.

A 'review' such as you suggest with multiple samples that really caught my eye was run by Asus, here:

http://www.hardocp.com/article/2013/06/01/intel_haswell_i74770k_ipc_overclocking_review/6#.VQ3eGPnF8xj

and can be looked to as to what you might run into as these components could all end in the same type of mixed results.

 

[Tradesman1]

This is one of the best DRAM related articles I've read. Great to see the issues of incompatibility by mixing sets are being emphasized on. The "simulation" test is a good measure for quantifying real world differences of different DRAM speeds and capacity.

A lot of work hardwork has been put in this, and it's totally worth for such an eye-opening article that this is. I fancy linking this to solve many related issues which are raised in the forums on a daily basis. Looking forward to more articles like this! Great work!

________________________

Thanks,

Am working on the followup now, looks like it will hit 29 common questions and fallacies about DDR3

 

[Ashu Malik]

Thanks you Sir.You always write the genuine stuff and before reading this article I never knew that putting two separate sets of sticks into a single motherboard can pose such a problem.






Erronous link removed - Tradesman1