DDR4 Frequency vs Latency

[Nicholas Peyton]

As you know higher frequency DDR4 is often attained at the sacrifice of latency (higher timings). The increased latency (timings) often completely outweigh any benefit to the increased frequency.

Companies continue to sell higher frequency (higher timing) RAM, knowing that people will be duped by the thought of "higher frequency = automatically faster". - is this really the case?

I happened across this formula that lets us calculate the answer:

CAS x 2000 / Speed(in MHz)
The lower the answer, the faster the RAM.

Examples:



Flare X (for AMD) F4-3200C14D-16GFX

DDR4-3200 (PC4-25600)
CL14-14-14-34
1.35 Volt
14 * 2000 / 3200 =8.75 <---- clear winner



Flare X (for AMD) F4-2400C16Q-64GFX

DDR4-2400 (PC4-19200)
CL16-16-16-39
1.2 Volt
16 * 2000 / 2400 = 13.32 <---- slower (as expected)




HOWEVER look at this 4133MHZ kit with absurd CL19 timings:

Trident Z F4-4133C19D-16GTZKWC

DDR4-4133 (PC4-33000)
CL19-19-19-39
1.35 Volt
19 * 2000 / 4133 = 9.19 <--- obviously still faster than the 2400 kit but actually slightly, slower than the 3200 kit


Any thoughts?

Are my findings/source flawed or is there validity in this?

Nick.

 

[sneekypeet]

Crucial memory did some whitepaper on speed versus latency. I am mobile right now, but essentially the higher latency with higher speed follows DDR3, and is not as off as one might initially assume.
http://www.crucial.com/usa/en/memory-performance-speed-latency

 

[Nicholas Peyton]

thanks ill check that out


Edit:

I looked at that table on the website you linked me, it seems my equation still has some use:

CAS * 2000 / Speed(in MHz)

If you take the last entry for example; the result is as follows:
18 x 2000 / 2666 = 13.50337


Or from the DDR4 2400 entry:
17 x 2000 / 2400 = 14.166666

So my original equation (1st post) is actually a calculation of "true latency", not true speed: (as I now learned)

At least I'll be able to use the equation to determine if my RAM upgrade is really an upgrade.

For example my current DDR3 2133 C12 kit (as I now learned) has a "true latency of": 11.25

If I upgraded to DDR4 3200 C16, my new true latency would be: 10. (16 x 2000 / 3200 = 10)

(So in reality I'm actually getting slightly lower "true" latency despite the higher timings AND the hugely increased speed).
It all makes perfect sense now ;-)

That link really did help!

 

[Totally]

UsHigher the frequency the more diminishing returns on tighter timings. The higher up you go on freq the less impact latency has.

DDR4-3200

@CL16

16 * 20 / 32 = 10

@CL19

19 * 20 / 32 = 11.875

DDR4-3000

@CL16

16*2 / 3 = 10.67

@CL19

19* 2 / 3 =12.67

 

[RejZoR]

What platform are you using? If it's X99, people say timings matter more than clock speed... Which kinda makes sense since you have quad memory channels. You already got tons of bandwidth...

 

[erixx]

I have been benching with my timings (everything else the same, including cache multiplier at auto) and the gains in AIDA Ram and Cache test are close to nothing moving from 16(-16-16) to 15 to 14 with primary speed of 3200Mhz for all. With 14 I got noticeable microstutter moving the mouse. So back to 15 and call it a day.

 

[Aenra]

Have been wondering about this since i started planning my current rig. While i too came across the aforementioned crucial URL, the fact of the matter is its findings are cherry-picked, as the data itself is cherry-picked. Data which (assuming one wished to extrapolate) isn't always as readily available as may be apparent. Least not to the average Joe (such as moi).
Add to that what one can find in places like overclock.net/f, where a close reading reveals that it's all basically bullshit outside of synthetic benchmarks? And you get where i have been since back then, lol..

Nowhere

As to bandwidth, that is way over-simplifying too, again wrong.. to put it kindly. You can fit 8 instead of 4 cars in a highway sure, but if the toll stop right ahead can only take two or four simultaneously, it all goes out the window. Speed always mattered the most and always will.
No offense, but you have a blog. For someone having a blog (ergo people read your stuff), you sure are.. liberal with your conclusions. Just my humble personal opinion, may well be mistaken.

So for now? Until i find someone who is i) knowledgeable, ii) mature (yes, i mean just that, because above a certain degree, nerdiness can be problematic) and iii) possessing numbers in the truckload so as to exclude the margin of error?

I stick to the empirical.
Empirical has told me that:
- for the little shit that make or break your day, timings come first.
(open explorer, open internet explorer, switch a tab, how fast thumbnails load if you have disabled saving them, etc.; like i said, petty stuff, but when the hours in front of the PC start adding up, they're what ultimately makes a difference towards that elusive "better" experience).
- for 'serious' stuff, such as heavy/intensive programs, speed comes noticeably ahead. Always.

I picked my poison and chose timings.

 

[agent_x007]

Timings only affect latency.
Basic math : 16 => 14, gives you drop of 12,5%, 18 => 16 only ~9%.
With memory latency in AIDA64 Cache&Mem benchmark of 50ns (on 18 timings), dropping all primary timings to 16 would net you a over 46ns result (in THEORY).
In practice timing scaling is REALLY hard to predict.
Basicly : You will see a gain in numbers, but how big will it be is anybodys guess.

On the other hand, frequency of RAM affects both bandwidth and latency (because timings are measured in cycles, and frequency decides how long that "cycle" is).
Other than that, bandwidth (the by-product of higher frequency) is required/VERY important for iGPU's.
So going after frequency is a reasonable goal to have on manufacturers side (they could produce chips that did 2000MHz CL7 on DDR3 side, so 4000MHz and 14 timings shouldn't be far from "golden goal" I have for memory).

 

[Aenra]

bandwidth (the by-product of higher frequency) and is required/VERY important

Correct me if i'm mistaken, but is that not in theory as well? Because in practice, we have precious few (if any) casual-oriented programs that benefit from very high bandwidth.
(had assumed we were talking exclusively about the practical)

 

[agent_x007]

Have you ever played on the same iGPU (APU), but with higher bandwidth available ?
You will see a difference, but don't trust me, trust him => LINK

Bandwidth is also needed just as much as latency, in high FPS game situation (those sweet sweet 120-180FPS we all crave for ).
More FPS = more data CPU is processing each second, which in turn requires more data to be delivered to CPU from RAM in each CPU/RAM cycle.

 

[Aenra]

You will see a difference, but don't trust me, trust him => LINK

You just lost all credibility arguments-wise, sorry. The "guy" you refer me to is the one giving wrong advice on so many things i have lost count. Latest of which that i do recall being how to delid..

Now as to games, 180FPS? You have a goal of generally/always having 180 FPS? Seriously? I mean to each their own, but..
If the basis for reaching a conclusion is what happens in games of all things, when at 180FPS of all possible scenarios.. whatever i guess. But we did say practical for a reason

Since however you stick to the empirical, i can tell you that with the rig in my sig, there's a maximum of 4, 4 (as an average, can be anything between 2 and7ish) FPS difference between 8, 16, 24, or 32 Gigs of RAM running quad.. counted in as many modern games as i own and quite a few i do 'not'.
So there goes your bandwidth in games argument ^^

edit: needless to say, with both channel and rank interleaving enabled, am not trying to cheat my own self.

 

[agent_x007]

You just lost all credibility arguments-wise, sorry.

So you don't like Linus - OK, I understand.
But I now "lost all credibility arguments", because of that... wow.
FYI : Just because you don't like someone, doesn't mean he's wrong in everything he does.

From logic perspective then :
GPU's like bandwidth more than latency (HUGE data to be processed by HUGE number of small "cores").
Add to that : iGPU's have to share bandwidth with CPU.
This isn't good for GPU's performance, and higher DRAM frequency helps direcly with that.

Because in practice, we have precious few (if any) casual-oriented programs that benefit from very high bandwidth.

Well, you wanted casual, you got it.

That 180FPS isn't for maximum/average.
It's for minimum in every game (stable FPS is, to me, more important than higher value itself).
That's why I say it's practical, I could have said 240Hz (top gaming monitors can show it), but I opted for 180FPS (stable).

EDIT : And what framerate you got on that X99 (average values of 2-8 can mean different things at 20-ish FPS, and 180-ish FPS) ?
If it's not 120-180FPS (minimum) - I'm sorry, but you don't have the data I want.

 

[Aenra]

Ain't about what you want though, don't you see? Because what you want is just in one aspect, in some extreme circumstances no one sane enough (or with healthy interests) should give a damn about in the first place. Or more to the point, should be expecting; there was distinct mention of the 'now', of the where we stand currently in terms of taking advantage.
Anyway, as i said above, to each their own and you're welcome your personal goals, but you are overreaching if you think the niche within a niche that you set is anything "concrete", or universal, to go by. If that makes us all casuals to you, be our guest.
(my reference to casual was literal btw, ergo what most use, ergo what most appreciate to extrapolate by)

To anyone else reading, i forgot to mention something even more important:
According to the theory we all like to believe, bandwidth = the more the merrier, ergo same scaling upwards right?

Wrong. I get the biggest benefit from adding 8 gigs to my 8 gigs, totalling 16. I get a significantly smaller benefit (close to negligible/margin of error) from adding another 8 to a total of 24. And i get usually nothing when adding yet another 8, to a grand total of 32.
Yes, the exact opposite. Diminishing returns to a point where i, (just me), can only theoretically concede that 'yes, more is better'. The actual benefit is still limited at 16gigs (see too close to negligible differences above).

edit: but we went off topic entirely; my apologies. I'm really interested in frequency vs latency so let us stick to that. Again, my sincere apologies.

 

[hat]

By your calculations, my 1333MHz DDR3 is faster than that DDR4 2400 you picked out (but 1-2 slower than the other kits).

I don't think this metric, by itself, means anything in the grand scheme of things.

 

[agent_x007]

By your calculations, my 1333MHz DDR3 is faster than that DDR4 2400 you picked out (but 1-2 slower than the other kits).

I don't think this metric, by itself, means anything in the grand scheme of things.

I guess OP didn't read the article which that table is used in : LINK (OP, ALWAYS provide source link if you are "borrowing" something from other peoples publications).
But back to main topic, article states :

Memory speeds were DDR4-2133 C15 and DDR3-1866 C9 respectively.

Which means DDR3 speed (based only on "true latency"), should be MILES ahead of DDR4 memory in latency.

Reality check for actual performance :

^and THAT is best case scenario for this set of DDR3 memory, FYI.

 

[BiggieShady]

diminishing returns on tighter timings.

... and on top of it, even if you reduce ram module latency by whopping 20%, it translates to only 4% reduction of latency on the memory controller ... talk about diminishing the diminished

 

[basco]

and plz dont forget that the second and 3rd timings mean often more on newer platforms since ddr4.
so its always hard to compare(like patriot has diff 2\3 timings then g.skill).
but you know now nick why on hwbot they shoot for 4000mhz with cl11 to 12.

 

[EarthDog]

Crucial memory did some whitepaper on speed versus latency. I am mobile right now, but essentially the higher latency with higher speed follows DDR3, and is not as off as one might initially assume.
http://www.crucial.com/usa/en/memory-performance-speed-latency

+1 no different really.

 

[Woomack]

You can count theoretical latency but it doesn't mean anything on new platforms. There are many other factors which are affecting performance like memory controller, memory ranks and many sub timings which are never in theoretical latency calculation.
On new platforms, higher frequency almost always wins with lower timings if we go for max performance. The best is balance but in general the best for desktop platforms are dual rank, high frequency memory kits.

 

[Enterprise24]

Z170 / Z270 frequency over latency.

If you grab b-die and overclock it. You will get frequency without sacrifice latency much

Actually 4133 19-19-19 will rekt 3200 14-14-14 anydays. Not speaking about theory but according to my own various benchmarks. And 4133 19-19-19 kit is very good for only 1.35V I am sure those kit can take 1.45V and still run stable at 16-16-16.

If you have knowledge about memory oc you can still grab 3200 14-14-14 since it is the same b-die as 4133 19-19-19 but slightly worst result when oc. My friend grab HOF 3200 CL14 and oc it to 3866 16-16-16 1.45V.

 

[hat]

I guess OP didn't read the article which that table is used in : LINK (OP, ALWAYS provide source link if you are "borrowing" something from other peoples publications).
But back to main topic, article states :
Which means DDR3 speed (based only on "true latency"), should be MILES ahead of DDR4 memory in latency.

Reality check for actual performance :

^and THAT is best case scenario for this set of DDR3 memory, FYI.

It looks like they're comparing some really good (not the best, but really good) DDR3 against some crap DDR4.

 

[Nicholas Peyton]

(OP, ALWAYS provide source link if you are "borrowing" something from other peoples publications).

You obviously never read my post properly.

Omg - I stated that I copied the table from a link that sneekypeet provided me (in the post directly above my own -- which had the link in it).. I also thanked him for providing the link lol.

See below:

Moreover; I only copied it to help better articulate what I thought I'd managed to 'grasp'!

Omg I've only been back here a day (after getting sick of listening to negativity before). And already I've been moaned at myself and witnessed two others on my thread arguing r.e something else too. (this being my 1st thread since I left).

Why are people here so "critical of others"?

If you've got nothing nice to say, don't say it at all. (thats my motto). Wish others could do the same.

That doesn't mean people can't disagree (that's not what I'm saying either -- even disagreement can be useful).

 

[Nicholas Peyton]

I've been looking further & happened across a RAM Overclocking "calculator" that lets you:

• Set the DDR speed "you want" (I.E. your target)
• Allows you to enter vital info regarding your sticks
• And now most importantly: actually calculates the correct (or best), 'but conservative' timings you need; to get that speed working on your sticks. (using the info you inputted)

If anyones interested in comparing results together (using the calculator) let me know.
Obviously all credit goes to the author, not me. But I could really use someone to bounce off with this. (pm me or reply here)

Thanks, Nick

 

[Caring1]

Welcome back, always interested in reading about your exploits.

 

[Aenra]

Welcome back, always interested in reading about your exploits.

My first thought was to (joke) and reply to him with something like: "I've heard latency improves when on LN2"
But as he's probably capable of actually trying it, lol

( no offense meant Nicholas, just joking ^^ )
(( although come to think of it, you already have a Jeesus chip. Imagine the potential, pairing it with a Jeesus RAM ))