RAM order of importance for: CL, TRCD, TRP and TRAS?

[nutzaalex]

Hi  

 

For a couple of weeks now I've been at it overclocking my ram. I am at the point where I've taken some steps back and am trying a different approach.

 

I can't seem to find an up to date guide on the OC importance order for memory timings, out of everything I read I understand that the order is:

 

CL

tRP

tRAS

tRCD

 

Is this accurate?

 

Also does DDR4 15-18-19-32 3600mhz look better than 17-20-20-38 3800mhz? I'm still testing the latter and currently found the lowest stable to be CL of 17, next up should I try to lower tRP first or tRCD?

 

On one hand I see Luke saying in a 3 year old video that in theory CL/ram speed is the important formula, so 15/1800 is better than 17/1900, issue is that I'm not even done lowering the 3800 OC timings and I notice better benchmarks on 3800's current stable timing values...

 

Also I got a voltage of 1.36 set > my msi B450 mortar max sets/reports it set to 1.38v so I assume that's a real setting and won't up the voltage since I wanna be safe, but I increased the timings a lot to become stable for now.

 

 

 

 

 

 

Extra details to my OC journey:

 

I was stable in memtest86 for nights/ more than 7-8h, past 8-9 back to back tests multiple times while working on the 3600 timings, made the mistake to go into secondary timings one by one, got it so weirdly unstable that not even 16-19-19-36 (original XMP is 3600 16-19-19-39) perhaps it had to do with memory training/ how motherboard set some secondary timings automatically while I was messing with some of them but not all since dram calc seems unreliable for me.

 

So after a couple bios sets/ resets and some more memtest86 from square 0 I went a night of testing the ram in memtest86 at XMP profile values, which passed (still there's a small buggy monster inside my brain that wispers I'm not 100% sure the ram's not broken after that wierd all of a sudden not stable by reverting back over and over again to previously deemed stable timings due to nights of memtest error free passes)

 

But RAM doesn't all of a sudden error out in memtest like first/second pass consistency then consistently pass multiple passes without errors?

 

 

As you can see I've got a lot of uncertainties on my mind... but the journey continues, if it's bad RAM at least I'll see it fail again now while tightening timings than have it fail in windows (which it never did since I'm doing this, I only modify then test in memtest86).

[Guest]

We are talking minimal differences.

 

What was the question?

 

The timings you are using for example are just terrible anyways. 

 

And what system would the memory inquire be for?? AMD? Intel??

 

Sorry was TLDR it all......

[nutzaalex]

41 minutes ago, ShrimpBrime said:

We are talking minimal differences.

Kinda agree but not quite, I do notice a minimal difference from xmp stock profile to ether my 3600 oc or 3800 wip one, I'm trying to figure which one is better by having the lowest timings for both because I like experimenting and it's not like I can go outside for example.

41 minutes ago, ShrimpBrime said:

What was the question?

TLDR: The question is in the title actually: RAM order of importance for: CL, TRCD, TRP and TRAS? 

 

long question: when generally trying to tighten primary ram timings, what is the order of importance/ which one to start with, second etc? by rule of which single timing has the most impact to lowering latency, confused because some depend on others.

 

I deduced so far that most impactful: CL, then tRP, tRAS, tRCD, is this accurate?

 

-logic behind the question: I'm gonna tighten timings anyways so why not start to lower to the max the most important values first?

41 minutes ago, ShrimpBrime said:

The timings you are using for example are just terrible anyways. 

 

15-18-19-32 3600 (my tRCD is split in 2: one being at 17T another 18T but for simplicity I wrote 18T) which better than 16-19-19-39 3600

 

17-20-20-38 3800mhz - this is wip, CL won't go lower, I'm in the process of lowering the rest.

 

41 minutes ago, ShrimpBrime said:

And what system would the memory inquire be for?? AMD? Intel??

 

I've got an amd ryzen 5 3600 4.3ghz all cores 1.25v

 

I'm at a safe voltage and that shows me a difference, not big, but once I'm done that difference will continue to exist, no more effort at that point, won't go into tightening secondary timings, I've seen the pain to attempt that and I would say it is at that point we're talking about imperceivable /minimal differences like you stated, I'd add the difference is only in benchmarks then.

Edited May 6, 2020 by nutzaalex

[Guest]

16 minutes ago, nutzaalex said:

Kinda agree but not quite, I do notice a minimal difference from xmp stock profile to ether my 3600 oc or 3800 wip one, I'm trying to figure which one is better by having the lowest timings for both because I like experimenting and it's not like I can go outside for example.

TLDR: The question is in the title actually: RAM order of importance for: CL, TRCD, TRP and TRAS? 

 

long question: when generally trying to tighten primary ram timings, what is the order of importance/ which one to start with, second etc? by rule of which single timing has the most impact to lowering latency, confused because some depend on others.

 

I deduced so far that most impactful: CL, then tRP, tRAS, tRCD, is this accurate?

 

-logic behind the question: I'm gonna tighten timings anyways so why not start to lower to the max the most important values first?

 

 

I've got an amd ryzen 5 3600 4.3ghz all cores 1.25v

 

I've also posted my timings so far, they may not look the best but I've mentioned I'm currently working on lowering the 3800mhz timings so those are wip, while the 3600 oc has lower timings than the xmp stock configuration, I'm at a safe voltage and that shows me a difference, not big, but once I'm done that difference will continue to exist, no more effort at that point, won't go into tightening secondary timings, I've seen the pain to attempt that and I would say it is at that point we're talking about imperceivable /minimal differences like you stated, I'd add the difference is only in benchmarks.

CL first, the rest to follow.

 

Goal = 

3600mt/s 16-16-16-16-26-46 1T

3800mt/s 18-18-18-18-28-52 1T

1.45v or less.

Trfc plays a role in performance. One of the many sub timings, but with most impact. 

Since you are not doing competitive benchmarking, you want stability over anything. 

 

Sometimes the SOC voltage needs an increase to help stability. 

If running odd CL latency, you want to disable gear down mode.

If overclocking, you might want to disable power down mode as well.

 

But definately try and smooth out your timings best as you can. 16-18-18-18 would be cool for example.

 

 

[nutzaalex]

16 minutes ago, ShrimpBrime said:

But definately try and smooth out your timings best as you can. 16-18-18-18 would be cool for example.

 

What benefit does that bring to have close timings between themselves over lowering them to the max? my xpm profile states 16-19-19-39, tRP won't go lower than 19 with 1.36v(1.38 v bios reports) since tRP won't go lower I ended up with 15-18-19-32 3600, same for tRCD won't go lower than 18T even with CL at 16. Since those timings are stable, I do pefer those over something like 16-19-19-x

 

 

Also keep in mind I got hynix-drj die, not samsung b or other overclocking monsters also don't want to go over 1.38v, so more v is out of the question for longevity concerns.

 

 

 

EDIT: I see now you've upped tRP up quite a bit in your recommandation, lemme digest

 

 

[nutzaalex]

24 minutes ago, ShrimpBrime said:

Sometimes the SOC voltage needs an increase to help stability. 

I upped it to 1.12..something, I've red 1.2 would be max, don't wanna up it more than it is.

 

24 minutes ago, ShrimpBrime said:

If running odd CL latency, you want to disable gear down mode.

If overclocking, you might want to disable power down mode as well.

All disabled

24 minutes ago, ShrimpBrime said:

3600mt/s 16-16-16-16-26-46 1T

3800mt/s 18-18-18-18-28-52 1T

the way they look 16/1800 is faster than 18/1900 this would mean 3800 is not even worth a try?

Also how did you come up with the values quite like this? I find it very intriguing since I understood by cl being the most important > a reason why it should be 15, not 16 if it can go one tick lower

[Guest]

6 minutes ago, nutzaalex said:

I upped it to 1.12..something, I've red 1.2 would be max, don't wanna up it more than it is.

 

All disabled

the way they look 16/1800 is faster than 18/1900 this would mean 3800 is not even worth a try?

Also how did you come up with the values quite like this? I find it very intriguing since I understood by cl being the most important > a reason why it should be 15, not 16

Well 15 is a 2133 Jedec timing. As you noticed this is the only time "gear down mode" disables it'self. 

AMD's memory controller doesn't typically like odd cas latencies and why it's standard most XMP profiles at an even number because I think Intel is the same way.

 

If you where going to cut nanoseconds, might as well aim for 2 of them. So no CL 15 isn't better on picky IMC, but CL 14 at 3600mt/s is doable with the right memory. My 2700x will do it, but with good b-die kit and some increased voltage that would be beyond your comfort zone. 

 

3600mt/s at CL 16-16-16 would pretty much be my suggestion for your daily use.

 

If you want accomplishment numbers, I have some of those as well. I'll just hint your pretty low on tweaking voltage. You need some scaling.

 

[Guest]

25 minutes ago, nutzaalex said:

What benefit does that bring to have close timings between themselves over lowering them to the max? my xpm profile states 16-19-19-39, tRP won't go lower than 19 with 1.36v(1.38 v bios reports) since tRP won't go lower I ended up with 15-18-19-32 3600, same for tRCD won't go lower than 18T even with CL at 16. Since those timings are stable, I do pefer those over something like 16-19-19-x

 

 

Also keep in mind I got hynix-drj die, not samsung b or other overclocking monsters also don't want to go over 1.38v, so more v is out of the question for longevity concerns.

 

 

 

EDIT: I see now you've upped tRP up quite a bit in your recommandation, lemme digest

 

 

Look many statements and questions while I was typing eh?!

All good.

 

Think of timings like speed limits on the road. 

speed up, slow down, speed up, slow down...... Just set up a cruising speed. It will give good performance. 

[nutzaalex]

1 minute ago, ShrimpBrime said:

Well 15 is a 2133 Jedec timing. As you noticed this is the only time "gear down mode" disables it'self. 

AMD's memory controller doesn't typically like odd cas latencies and why it's standard most XMP profiles at an even number because I think Intel is the same way.

 

If you where going to cut nanoseconds, might as well aim for 2 of them. So no CL 15 isn't better on picky IMC, but CL 14 at 3600mt/s is doable with the right memory. My 2700x will do it, but with good b-die kit and some increased voltage that would be beyond your comfort zone. 

 

3600mt/s at CL 16-16-16 would pretty much my suggestion for your daily use.

 

If you want accomplishment numbers, I have some of those as well. I'll just hint your pretty low on tweaking voltage. You need some scaling.

 

aha!

 

This is what I was looking for, parts of this puzzle fit now and it would explain possibly why sometimes I appeared to have issues with 15 that I can't explain or reproduce.

 

 

So my game now is to go 3600 16-16-16 -lowest it can go- lowest it can go, keeping both even numbers, do the same for 3800, though I doubt it I can hit 18-18-18 without more voltage if possible at all, maybe I'll use 1.4v to see and compare which one is faster and leave it as is. Though people reported for ddr4 not 100% safe to go 1.4 or >

 

 

I am dissapointed though that most likely I won't end up using 3800 one while I was happy I could oc infinity fabric to 1900

 

[nutzaalex]

2 minutes ago, ShrimpBrime said:

Look many statements and questions while I was typing eh?!

All good.

 

 

All is now less cloudy in my head regarding this, thanks, curious if I can find those even and equal timings you're talking about

[Guest]

Quote

Also keep in mind I got hynix-drj die, not samsung b or other overclocking monsters also don't want to go over 1.38v, so more v is out of the question for longevity concerns.

Well, you are overclocking and tweaking on stock or really near stock XMP voltage. 

Try 1.450v maybe.

 

_________

So when I am pushing numbers, let's say I'm between 1.50-1.60v CL14-14-14 3600mt/s on b-die.

The deal with the all mighty b-die is simply it scales very well with increased voltage. 

Some Hynix and Micron will scale well also. Just depends how high tier memory you purchase.

 

I use active cooling on my memory by the way. The do not rely on the cpu fan for cooling. The system is typically run on an open bench as well.

[nutzaalex]

6 minutes ago, ShrimpBrime said:

I use active cooling on my memory by the way. The do not rely on the cpu fan for cooling. The system is typically run on an open bench as well.

I got noctua fans pushing in air from pretty close, don't have active cooling though, my ram also dont have a temp diode on them so I wouldn;t know if 1.45v is too much, that's why I wouldn't go there, perhaps I'll put a probe sometime but till then I wouldn't try it.

 

my ram: F4-3600C16D-32GVKC

G.Skill Ripjaws V DDR4-3600MHz CL16-19-19-39 1.35V

 

if this means anything, stock voltage for xmp is 1.35

[Guest]

8 minutes ago, nutzaalex said:

I got noctua fans pushing in air from pretty close, don't have active cooling though, my ram also dont have a temp diode on them so I wouldn;t know if 1.45v is too much, that's why I wouldn't go there, perhaps I'll put a probe sometime but till then I wouldn't try it.

 

my ram: F4-3600C16D-32GVKC

G.Skill Ripjaws V DDR4-3600MHz CL16-19-19-39 1.35V

 

if this means anything, stock voltage for xmp is 1.35

I understand the concern. 

[nutzaalex]

so just for testing: I did 1.45v to see at least if it works without planning to stress it yet, 16-16-16-26-48 was out of the question, not even boot, will try the 3800 one but I already had low confidence it those timings getting so low after all my tests

 

on 3800 timings: 

1 hour ago, nutzaalex said:

3800mt/s 18-18-18-18-28-52 1T

tRP only goes to max 27, no 28 possible, 18-18-18-18-27-52 boots but not stable, so double nogo even with high voltage

 

so question becomes if possible/ which one is better at 16-18-18-18.. 3600 vs 18-20-20-20-... 3800

 

Anyways don't want to bother you anymore, thanks for the help!

 

I'll continue trying out formats like this if there isn't any one stable though, for example 16-18-18 one, there's no other chance than all over the place timings

[Guest]

2 hours ago, nutzaalex said:

so just for testing: I did 1.45v to see at least if it works without planning to stress it yet, 16-16-16-26-48 was out of the question, not even boot, will try the 3800 one but I already had low confidence it those timings getting so low after all my tests

 

on 3800 timings: 

tRP only goes to max 27, no 28 possible, 18-18-18-18-27-52 boots but not stable, so double nogo even with high voltage

 

so question becomes if possible/ which one is better at 16-18-18-18.. 3600 vs 18-20-20-20-... 3800

 

Anyways don't want to bother you anymore, thanks for the help!

 

I'll continue trying out formats like this if there isn't any one stable though, for example 16-18-18 one, there's no other chance than all over the place timings

At least you are putting in some good effort. That's more than most people. 

I always say it. Tweak and test. Tweak and test. It's the only way.

Still play with mine even though I already know or have presets that give me go go.

Am fortunate as well to be blessed with some very decent memory. 

 

I do have experience with Hynix sticks. Have a couple sets of Corsair 3000 LPX and LED. They do about 3466mhz at their very best big 1.60v. In short, does not scale well.

 

---

 

OK so difference between 3600 and 3800 with timing differences CL 16 vs 18. Well the latency will be similar. The read write and copy will be faster on the 3800 from clock speed. Also be aware that the cpu frequency makes a difference. Always test with the Cpu the same speed for all the memory testing. That's important. A lower cpu frequency will yield higher latency and slower reads and writes. But I can also tell you, you fight that memory controller. 

 

Really wish you had some eye candy so I can better see what you are looking at. My imagination can run rampit and I start thinking your Ram can clock like mine. lol.

 

 

[2dfx]

CL is always the first important value to consider when performing RAM OCing.

 

Column Access Strobe Latency or CL for short, is the delay between when the memory controller requests a readback from the memory, and when the data in memory is actually available to the controller.

 

CL isn't bound strictly to clock speed.  For DDR4, a CL of 14 at 3600 has a first word speed of 7.784ns but a CL of 14 at 3866 7.252ns.  There are calculations you can do to get a rough idea of raw clock speed vs latency.

[nutzaalex]

19 hours ago, 2dfx said:

CL isn't bound strictly to clock speed.  For DDR4, a CL of 14 at 3600 has a first word speed of 7.784ns but a CL of 14 at 3866 7.252ns.  There are calculations you can do to get a rough idea of raw clock speed vs latency.

Thanks for the answer, I think I do have an idea on the rough calculation that is CL/ram speed, in my latest case/progress CL 16 3800 so 16/1900 is the calculation also based on what Luke said in a video, this gives a rough estimate of the latency.

 

I am interested in the rest of the values as well, like what comes after CL as being most important? how about the third value? etc.

 

My idea formed on reading the internets on that order is: 

1 CL

2 tRP

3 tRAS

4 tRCD

 

@ShrimpBrime pointed out that the order of importance as I understood it is the order of in which the latencies are declared so tRCD second, tRP and lastly tRAS.

 

Also CL must be even, the others don't look they need to be even though.

 

Another thing is that it seems better to try and have close numbers between each other, like 16-16-16 not 15-17-18 etc.

 

 

 

 

 

20 hours ago, ShrimpBrime said:

Am fortunate as well to be blessed with some very decent memory. 

haha , I also feel the same way with the ryzen 5 3600 I got, since it goes 4.3 all cores with 1.25V, I find it a decent win in silicon lottery, I also ditched 4.35ghz just because it would need @1.28v or something similar. and figured 4.3 not so stressed is good enough.

 

20 hours ago, ShrimpBrime said:

I do have experience with Hynix sticks. Have a couple sets of Corsair 3000 LPX and LED. They do about 3466mhz at their very best big 1.60v. In short, does not scale well.

Spoiler

 

Well, so I have F4-3600C16D-32GVKC G.Skill Ripjaws V DDR4-3600MHz CL16-19-19-39 1.35V XMP - hynix djr, I think I mentioned that before, but gonna mention again to have it all in one post.

 

20 hours ago, ShrimpBrime said:

Also be aware that the cpu frequency makes a difference. Always test with the Cpu the same speed for all the memory testing. That's important. A lower cpu frequency will yield higher latency and slower reads and writes. But I can also tell you, you fight that memory controller. 

First thing I tuned my CPU, it was always 4.3 all cores all the time, same voltage, so I consider this a stable constant. I also upped a bit the CPU SOC voltage, then lowered it to lowest stable voltage, since 1900 infinity fabric also works stably I'd say I found the lowest stable values there.

 

20 hours ago, ShrimpBrime said:

Really wish you had some eye candy so I can better see what you are looking at. My imagination can run rampit and I start thinking your Ram can clock like mine. lol.

Ah, I'm glad and excited to show you my build:

Spoiler

This is the inside of my pc, I've hooked a basic Celsius temp meter, in the ram's fins, since like I said there's no temp measuring on the ram, it showed me that at 1.40V set, aka 1.42v reported by bios the ram cooler doesn't go past 41 degrees, I know the temp on the actual chip is hotter than that, but I got no way to measure it better than this, would you say keeping 1.42v max operating voltage seems ok? also keep in mind the cpu cooler is gigantic (under 5mm between it and the side panel ) the cpu cooler fan rests on top of the ram sadly, but there's the other two fans that blow air on the ram from a sata ssd's distance as seen in the photos, so there is airflow there.

 

Also will get a corsair rm750 psu instead of what I have now in a week, and will make a build using this 650 psu for my dad. I expect that a higher psu won't cause any unforseen instability, if anything it should be even more stable..

Spoiler

 

There are three cuts in the ram heatsink, as you can see, I've stuck the probe inside one of them to read the temp.

Spoiler

 

Since we last talked, @ShrimpBrime I said already I tested 16-16-16-x 3600 and it is not stable as well as 18-18-18-x 3800 and it is not stable.

 

Now my thought process is: stay at 1.42V, once I am done finding the lowest stable timings, I'll try to lower voltage as much as I can as the last step (If I won't be able to at all, I'll test for 1.4v the lowest latencies I think)

 

I am trying to bring CL lowest, then tune the rest, lastly, tune tRAS.

 

I have bumped tRAS a lot, just so it's not a factor while identifying the lowest primary timings then deal with finding lowest tRAS.

 

Since my ram cant go low latency that much I ditched 3600 trials entirely.

 

I have ran last night a memtest like 10h, over 7 passes ok with: 16-20-20-20-40 3800 - so I deem this stable.

 

Currently I am testing 16-19-19-20-48 3800 - if ok will try 16-19-19-19-x if it goes well, then lowering tRAS to the min stable.

It is not possible to go 18-18-18 so I assumed it would be better to try CL 16 and lowest timings, and so far it does look like it's a win

 

I will share later how my bios looks, ram test, full system info because currently I'm running the stability test for 16-19-19-20-x and wouldn't want to stop a good 4h continuous progress

 

[nutzaalex]

18 hours ago, ShrimpBrime said:

Really wish you had some eye candy

https://imgur.com/a/eGkbX14

 

There you go, idk if it's eye candy though cuz its not low timings, but hey, its something to gain more speed and decent latency more stable than before

[Guest]

Holding off reply for 2 reasons. 

1 the wall of text to carefully read

2 Interested in @2dfx comments and thoughts. He seems more knowledge-able than myself. I dont typically take the time to calculate latency on the Cas latency timing, I thought it interesting he shared it.

[nutzaalex]

Ok well I ran memtest for 20h a total of almost 13 passes at 16-19-19-20-35 3800 1.37v set (1.39 bios reported) I think these are the best timings

 

Kinda itching to do tRC and tRFC cause why not, any suggestions for numbers/importance order?

 

https://imgur.com/a/WQvJ59M

[Guest]

Well whatever they are on auto, takr them off 1 at a time and lower the timing a bit.

Trfc do maybe 10 tics off at a time. This will help performance a little.

[nutzaalex]

Thanks for all the help! I'm on my way to lowering those as well, it will take a couple of days I guess

[Guest]

20 minutes ago, nutzaalex said:

Thanks for all the help! I'm on my way to lowering those as well, it will take a couple of days I guess

Cool! You're welcome.

[Islam Ghunym]

Do not punch me because I reviving an old thread. I came here by coincidence. When people use google or something and they end up here they need the information so I will write them:

Importance exactly goes this way:

 

1- Frequency + loose timings at high SA/IO or SOC voltage + proper impedance to run highest frequency.

 

2- CR >> tRRD >> tFAW >> tRFC >> tCl >> tRCD >> tRP >> tWR >> tRTP >> tWTR >> tRDRDSCL >> tWRWRSCL >> tWRRD >> tCWL >> tRAS > tRC >> fast safe run over other timings >> lastly longer tREFI

 

This order is by what you should tune first not by which gives higher performance. The reason is that timings effect each others and have relations. That's why you will find tREFI at the end while it gives a significant performance more than majority of the previous ones. loosening tREFI at the begining can prevent other timings from going as tight as it should. the same story for tWR which barely gives any performance, but it is tightened first because it allow other timings to go even lower (ex:tRAS) if it was tightened first.

 

By the way your XMP was 3600 and you are trying to do 3800. Do you call that an overclock seriously.... Such a a ram stick should do at least 4400MHz