The Speed and Latency question

[MayhemIam007]

Good day folks.

So I've read a couple of articles concerning memory speed, latency and the consensus seems to be, that if you are only going to use your PC for gaming (occasional Visual Studio Dev work) then one would not see an increase in Average FPS when going above 1600MHz speed.

So sticking with DDR3 1600MHz seems to be the consensus all round, if you are only gaming with said PC.

There have been reports that BF4 does get a bit of a boost in the max FPS department but not in the Average FPS (so it wouldn't make sense to go with faster than 1600MHz, also I'm not into online multiplayer games that much).

Q1
So the question becomes, when will games see an respectable boost in Average FPS when pushing one's memory higher?

Q2
What are the determining factors regarding a higher Average FPS count when playing games?

Q3
Why is there not an Average FPS increase when gaming on higher frequency memory.

PS, I know that when increasing one's Frequency your timings also diminishes.

Thank you for your time!

EDIT:
Sorry for not being a little bit more specific. I am talking hardware wise on Intel CPUs (intel i5), and running a beefy GPU (gtx970).

Q4
What would be a good comprimise between your timings and speed? Are we talking about 1866MHz CAS9, 2133MHz CAS9 or 2400MHz CAS 10/11

 

[iron8orn]

AMD recommends 1866mhz and while Intel recommends 1600mhz I have seen many say that Haswell runs better with faster and it's pretty common knowledge that apu's need 2133mhz+ to perform best.
You can get fast ram with good timing, just gotta pay for it.

 

[Outlander_04]

1/ there was an article here on Tomshardware comparing RAM frequency and gaming performance . A few games did better with faster RAM , but usually not cost effectively

2/ money . Cubic money usually

3/ the software of the game engine cant use the higher memory bandwidth available , or the higher latencies that go with higher frequency offset any performance increase

 

[iron8orn]

Makes since. Looking at newegg for the best.

Looks like:
cas 7 @ 1600mhz
cas 8 @ 2133mhz
cas 9 @ 2400mhz
cas 10 @ 2600mhz
cas 11 @ 2800mhz

cas 8 @ 2133mhz the best? http://www.newegg.com/Product/Product.aspx?Item=N82E16820233661 $142

 

[mdocod]

Effect on performance scaling, from most to least in most applications/conditions:

1. Channel interleave
2. Rank interleave
3. Speed
4. Timings

Channel interleave has a significant impact on bandwidth (doubles bandwidth). Performance scaling from nill to 50% or greater can be observed from channel interleave. (Greatest impact is on iGPU performance, where ~50% or more performance scaling is often observed). In traditional workloads, 10% or better performance scaling is not uncommon.

Rank interleave has a significant impact on access latency (halves it in some cases), which can manifest as improved bandwidth and/or computation performance in memory intensive workloads that are sensitive to access latency. A ~10% performance improvement has been observed in many enterprise environments with rank interleave enabled, and I have personally tested for this in more traditional desktop workloads and seen similar results in some cases. Rank interleave also effects iGPU performance by a margin greater than a full step up the memory multiplier ladder. (IE: rank interleaved 1866 is faster than 2133 speeds without rank interleave). Generally rank interleave provides more performance scaling than speeds and timings for traditional compute workloads. Raw speed can produce more performance scaling than rank interleave for powerful iGPUs.

Speed effects bandwidth and effective latency. Performance scaling with speed depends on the workload. Most compute workloads scale poorly with speed (less than 5% performance increase from a 50% increase in memory speed is common). iGPU performance can scale pretty good with speed. A 25% increase in performance from a 50% clock speed increase is pretty common.

Timings have an impact on latency and effective bandwidth. Performance scaling with tighter timings is generally very slim (<5%). Performance scaling on iGPUs from timings is generally negligible or non-existent.

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

All consumer/gaming/value/performance non-ECC DDR3 DIMMs configured to 8GB of capacity are dual rank dimms, so if you're buying a 2x8GB kit, you are already taking care of the 2 more important influential factors for memory performance (channel and rank interleave). If you're buying 4GB dimms, some are single rank, and some are dual rank. Most manufactures do NOT list rank or component configurations on "consumer" class DIMM specification pages. One of the last refuges for 4GB DIMMs configured dual rank, that I am aware of, is the crucial ballistix series. Outside of that it's hit and miss. Kingston and G.Skill are selling mostly single rank 4GB DIMMs, so I tend to avoid recommending them unless the price advantage is significant enough to outweigh performance loss (up to ~10%).

When buying 8GB performance dimms, if you are trying to eek out the best performance for the money, make sure to include consideration of the voltage required for the rated speeds and timings. Being able to run at 2133-10 isn't particularly special if it takes 1.65V to do it. Lots of 1600-9-9-9@1.5V kits will overclock to 2133-10 speeds with ~1.65-1.70V anyway. A kit sold as 1600-8-8-8@1.35V may be a better value than a kit sold as 2400-11-13-13@1.65V, as the low voltage kit can likely be tuned to a similar result.

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

On a final note: In my experience, The AM3 and AM3+ platform responds to the balance of speed/timings differently than 1150 and 1155 platforms. Rank interleave and timings are more influential on the AMD platforms than on Intel platforms in my testing and research, while channel interleave and speeds are more influential on Intel platforms, I am not sure why this is (I have some theories).

 

[iamlegend]

I am not well versed into RAM speed and timings but IMO the main usage of this 1600Mhz and 2400Mhz is for intel and AMD`s iGPU.

Stronger iGPU (AMD) recommends 2400Mhz of RAM to fully utilize the iGPU which means, your system will require higher memory to process both compute and graphical tasks.

While on the intel`s side, the recommended RAM is 1600Mhz which also comes with weaker iGPU, same as to say with dGPU which also has it`s own Memory.

 

[iron8orn]

Thank you for such a detailed response mdocod.

 

[MayhemIam007]

@mdocod

That was a legendary response! You rock dude!

 

[MayhemIam007]

@mdocod

Seeing that you seem to know your stuff, and with your answer in mind, I have 2 kits to choose from. Which one would you reccomend? As these 2 kits sits somewhere in between your recommendations.

Option 1
G.SKILL RipjawsX, 8GB Kit(4 Gig * 2), DDR3-2133, 1.5V, CL9 (F3-2133C9D-8GXL)

Option 2
Kingston hyper-x Savage, 16Gb kit (8Gb x2), ddr3-1600 , CL9 , 1.5v (HX316C9SRK2/16)

Again thank you for your valued input

 

[mdocod]

Hi Mayhemlam007,

Of those 2 kits, I would take the HX316C9SRK2/16 for 3 reasons.
1. twice the capacity.
2. dual rank dimms instead of single rank dimms (which is more useful for performance than the higher speed of the ripjaw kit).
3. lower profile heat-spreader for better compatibility with CPU HSFs.

I actually own that ripjaw X kit. Bought it about a year ago for an APU system. I was exploring performance scaling of different memory configurations and speeds, slipped that kit into a system expecting an increase in bandwidth benchmarks etc, and observed a LOSS in performance compared to another lower speed kit (ballistix sport). It was that peculiar discovery that lead me down the path of trying to figure out what's what with system memory. That was when I discovered the dirty secret going on with consumer DDR3 4GB DIMMs. Most of these 2133-2400 speed kits with 4GB DIMMs are configured single rank and won't perform any better than a 1600 speed kit configured dual rank. So many of these "high speed" kits being sold as performance class memory are actually highly misleading.

 

[MayhemIam007]

@mdocod

Hi mdocod

Man oh man. That is most startling news indeed, the scaly, tricksy manufacturers! This is very big news. Maybe Toms's should do an article on this? To let more folks know about this bait and switch trick of theirs. Look I am not dumb when it comes to PC hardware, but I am no expert, but manufacturers always tout their low Latency and High speed dimms from the roof tops as the be all and end all to your problems.

Have you tried sending Toms any of this info, so that they can maybe do an article on this?

Again thank you for your input and most precious info regarding this.

 

[mdocod]

Hello Mayhemlam,

I haven't made any effort beyond sharing the rank interleave issue in threads where relevant. The problem is that people are in love with brands like G.Skill, and don't want to entertain a thought of negativity towards them, especially when the issue can be dealt with by simply pretending it doesn't exist, so the issue is mostly met with disinterest and it doesn't soak in with anyone.

I have not contacted TomsHardware about the possibility of an article on the issue. I don't know what level of interest they would have in it, but there is one thing I would like to see manufacturers start doing, and that is list the component/rank configuration of every dimm they sell. Currently Kingston is the ONLY brand that does this. I don't know if such an article would persuade or dissuade a positive change with other brands.

 

[MayhemIam007]

Hi mdocod.

I can tell you 1 thing for certain. Let the facts speak for them self. I for one am open minded towards many a things in life, and one of those is knowledge (probably the biggest I'd say). It would be interesting to see if Toms shows any interest in this. One thing I hate is when some one thinks he can pull a fast one over my eyes, and the only way to counter that is with knowledge.

So I say give them the raw data, compile it, and send them a email. Cause in my opinion people with brand loyalty are like lambs for the slaughter house. When the facts and data are there then you need to seriously rethink your brand loyalty and if you don't want to then be a simple minded fool who can't think for himself. It's a choice at the end of the day. Go forth and spread the good news I say my man!

 

[MayhemIam007]

Also a quicky.

I've read on forums that your memory chips on the Dimms them self makes a huge difference, when it comes to overclocking said kits and that it's hard to find out what ICs will be on your sticks (like Hynix vs Samsung, where Samsung being the better option on overclocking). Now again, when reading reviews and such I have never ever read that from the person writing the article, and to me that would be a big deal as well.

On top of that, is it single sided or double sided memory on the sticks (where double side provides superior speed), but this I read on an article while I was looking through reviews on the sticks.

 

[mdocod]

The specific chips used to assemble the DIMM are indeed a factor in what sort of speeds and timings they will run at without errors, and is also a factor in power dissipation, voltage requirements, etc. Variations in fabrications used effect the these characteristics. Late generation DDR3 fabrication is significantly smaller and dissipates significantly less power than early generation DDR3, and also clocks higher. Interestingly enough, it would appear that in some cases, the same fabrication technology that is being used for this late gen DDR3, is being used for the first generation of DDR4. Samsung appears to have 20nm ICs available for both memory specifications.

Achieving really high speeds and really tight timings, like most extreme overclocking, does not make or break performance. Most modern DDR3 DIMMs will be able to run in the 1866-2133+ range pretty well if you find voltage and timings that suit the chips used on the DIMM, and most workloads will show very little (if any) performance scaling beyond this range anyway (with the notable exception being iGPU performance on APUs, which in some cases scales pretty well to 2400+ speeds). The memory performance tuning enthusiast crowd that is comparing the overclocking results of different memory chips are mostly just interested in milestones for the novelty of it. If you are enthusiastic about participating in the novelty of tinkering and performance tuning (or even competitive overclocking), then you may appreciate a memory kit like a G.Skill TridentX kit or other "enthusiast" class memory made from chips known to be stronger overclockers.

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

Unfortunately, single or double sided does not always tell us anything about the rank/chip configuration of a DIMM.

Memory chips are available in 4 - 16 bit widths. Each channel of a DDR3 bus is 64 bits wide (or 72bit for ECC memory). Thus, it's possible to assemble a single rank DIMM with only half of one side of the DIMM populated (4 x 16 bit chips). or with both sides populated (16 x 4 bit chips), or with one side populated (8 x 8 bit chips). It's also possible to assemble a dual rank DIMM with only one side populated (8 X 16 bit chips), or with both sides populated (16 x 8 bit chips).

Most 8GB consumer DIMMs appear to be configured with 16 X 8 bit chips, thus, they are double sided, with 8 components per side, dual rank. Most 4GB consumer DIMMs used to be configured this way when DDR3 was younger, but as the price of the larger capacity chips came down, more and more manufactures switched to using the same components found on 8GB DIMMs, on the 4GB DIMMs, only half as many of them (a single rank instead of 2).

It is possible to make a dual rank single sided 4GB DIMM using "dual ranked" (16 bit) components. For example, the MT41K512M16 from Micron could be used to build such a DIMM.