Tertiary timings on Ryzen

[Glaceon]

First off, specs:
Board: ROG Strix B550-F
CPU: Ryzen 9 3900X
RAM: 4x 8GB G.Skill Trident Z Neo 3600 CL16-16-16-36 1.35v (F4-3600C16Q-32GTZN) B-Die

So just got done with quite the RAM tightening session today. After a bit of tinkering (and some CMOS resets), I found that the current timings are stable:

Compared to the original timings I had:

Geardown cannot be disabled without some miracle, I'm done trying to get GDM off and 1T to be stable.

However I have a question about these timings:

Are these at all important or make a big difference on Ryzen, on top of the tightening I've already done? Getting these to be stable at all is a pain in my bum, would rather leave them alone.

This is my AIDA64 benchmark with all the timings I did manage to tighten:

On XMP, my latency was 72-74 ns.

Should I quit while I'm ahead?

In case you haven't caught on, whole point of this post is to know if those tertiary timings I marked above make any significant difference on Ryzen.

Update: managed to get tRDRDSCL and tWRWRSCL to 4, tWR to 10, and lowered tWTRL & tRTP to 8. Testing for stability now.

Another update: Seems like messing with either tCWL, tRDWR and tWRRD makes my system unable to POST. Might just ignore these. Am I fine to ignore these considering everything else I've done?

 

[dont whant to set it"']

Your screenshot shows something that is more than alright, I'd try make it stable if it is not already and by gones be gone.
Do not go by xmp more than primary timings as base ref. , Trfc is and should be different.

 

[TheLostSwede]

I would say that they have little to no impact on overall latency.
tRFC seems to be the most important one beyond tCL.
Considering you're "only" at 3600MHz, maybe try hitting 3733 or 3800MHz 1:1 instead of worrying about the minor timings.
I'm on Hynix CJR chips and this is what I get with four modules at 3800MHz.
Keep in mind I can't tighten my timings as much as you have.

 

[Glaceon]

Eh, I'm happy with 3600, last time I tried 3733 I had difficulties. Was on a Gigabyte board but still. I just want the best I can get @ 3600 CL16.

Here is the end result, after managing to fine tune the last timings:

Passed 2 passes in memtestx86 and stable in TestMem5 with anta777's extreme config.

Seems like touching tCWL completely kills any ability to POST.

Are the timings under tRDRDSC important at all tho?

 

[tabascosauz]

As long as RDRDSCL and WRWRSCL 4 or below, the others below don't matter a whole lot. There's *some* merit to CWL, RTP and RDWR, but I always just set 16/12/8 and forget (for CL16 obviously). A lot of the obscure ones like CKE, best practice is just to not think you're smarter than the board is. The board trains those the way it needs to to get you to POST.

Even for the usual secondary suspects (RRDS/RRDL/FAW/WTRS/WTRL/WR) I don't care for going tighter than 4/6/16 4/12/12. Regardless of what the diehards say. I've benched 4/4/16 4/10/10, all it does it require more voltage for literally the same or even worse performance.

And this is B-die. The single largest performance differentiator outside of tCL is tRFC. You need to be aiming at 160ns at the very least, if not lower at any speed.

320 @ 3600 is high. My 4Gb E-die did 325 @ 3600.

Refer to the Reous table for what ns is at what speed



Correction: I run 280 @ 4000, so 140ns. It's getting late. My B-die is average-to-shitty.

If you want the pros' opinions, I'm sure buildzoid talks about some of these tertiaries in one of his many B-die videos, and luumi has probably videos on them as well.

 

[TheLostSwede]

As long as RDRDSCL and WRWRSCL 4 or below, the others below don't matter a whole lot. There's *some* merit to CWL, RTP and RDWR, but I always just set 16/12/8 and forget (for CL16 obviously). A lot of the obscure ones like CKE, best practice is just to not think you're smarter than the board is. The board trains those the way it needs to to get you to POST.

Even for the usual secondary suspects (RRDS/RRDL/FAW/WTRS/WTRL/WR) I don't care for going tighter than 4/6/16 4/12/12. Regardless of what the diehards say. I've benched 4/4/16 4/10/10, all it does it require more voltage for literally the same or even worse performance.

And this is B-die. The single largest performance differentiator outside of tCL is tRFC. You need to be aiming at 160ns at the very least, if not lower at any speed.

320 @ 3600 is high. My 4Gb E-die did 325 @ 3600.

Refer to the Reous table for what ns is at what speed

View attachment 180702

Correction: I run 280 @ 4000, so 140ns. It's getting late. My B-die is average-to-shitty.

If you want the pros' opinions, I'm sure buildzoid talks about some of these tertiaries in one of his many B-die videos, and luumi has probably videos on them as well.

Interesting table. Sure explains some of the difference between different types of memory.

 

[Glaceon]

Tested 160ns and I had a black screen on cold boot. 140ns and no post at all. 180ns best I can do.

Should I attempt trfc 288 @ 1.4v?

Update: set tRFC/2/4 to 288/214/132, DRAM voltage to 1.4v.

and then shut off, and unplugged the PC.

5 minutes later, plugged it back in. Guess what? It boots instantly. Doesn't even sit on DRAM light anymore.

I hope this isn't just me getting lucky and that it stays this way.

 

[oobymach]

I run mine @ 160ns trfc but I'm only running 16gb, if 160ns is unstable try 175ns which is
trfc 315
trfc2 234
trfc4 144

Anything less than 160ns you might run into stability issues, and 140ns is about the lowest you can go with 2 sticks of ram. I use 1usmus ryzen ram calculator for ram timings.

DRAM Calculator for Ryzen v1.7.3 Download

DRAM Calculator for Ryzen helps with overclocking your memory on the AMD Ryzen platform. It suggests stable memory timing sets optimized for your m

www.techpowerup.com

 

[tabascosauz]

Tested 160ns and I had a black screen on cold boot. 140ns and no post at all. 180ns best I can do.

Should I attempt trfc 288 @ 1.4v?

Update: set tRFC/2/4 to 288/214/132, DRAM voltage to 1.4v.

and then shut off, and unplugged the PC.

5 minutes later, plugged it back in. Guess what? It boots instantly. Doesn't even sit on DRAM light anymore.

I hope this isn't just me getting lucky and that it stays this way.

tRFC is voltage dependent, and daily voltage up to 1.5V depends on what you can run without exceeding 50C on the RAM, with the heat of a GPU dumped on the RAM during games.

tRFC below about 160ns, if doable, you probably will not be able to sustain with the same voltage you were running at 180 or 190.

I'm not exactly sure why any of you are manually setting tRFC2/4, that's literally not a "thing". Just set tRFC and let the board do its job.

If it's not POSTing reliably then back off a bit on the tRFC ~10ns and test again. If anything doesn't POST reliably, it's as good an indicator as anything HCI or TM5 is going to be spitting errors.

 

[oobymach]

I'm not exactly sure why any of you are manually setting tRFC2/4, that's literally not a "thing". Just set tRFC and let the board do its job.

Manually setting all 3 timings reduces latency and increases bandwidth (read speed), it literally is a "thing". You can use 1usmus ryzen ram calculators additional calculators tab to find out what you need to set them at for what speed in nanoseconds you choose. Setting the first trfc sets all 3 to the same setting and you lose speed and bandwidth, you can verify with a benchmark like aida64 exactly how much you lose or gain by simply setting a few timings. And you can push your ram to its limits but 160-175ns is the butter zone.

Set the trfc manually if you want to reduce latency, if you let the board "do its job" it will set a trfc of 350ns which is like 600odd trfc because that's what it does when its not told what speed to run at.

In the end it's your computer not mine, I don't care how bad your latency or performance is.

 

[thesmokingman]

Manually setting all 3 timings reduces latency, it literally is a "thing". You can use 1usmus ryzen ram calculators additional calculators tab to find out what you need to set them at for what speed in nanoseconds you choose. Setting the first trfc sets all 3 to the same setting and loses you speed and bandwidth.

Set the trfc manually if you want to reduce latency, if you let the board "do it's thing" it will set a trfc of 350ns which is like 600odd trfc. I don't care what you do with your computer it's not mine I don't care how bad your latency is.

Huh?

The sub trfc are AUTO calculated from the main trfc setting. Your timings are all over the place. Those dimms can do cas 14, yet you have em set to 16. Then you set trfc to 288 which is cas14 range and run 1.4v. Like I don't get where you're going with these timings, much mismatched.

 

[TheLostSwede]

Manually setting all 3 timings reduces latency, it literally is a "thing". You can use 1usmus ryzen ram calculators additional calculators tab to find out what you need to set them at for what speed in nanoseconds you choose. Setting the first trfc sets all 3 to the same setting and you lose speed and bandwidth, you can verify with a benchmark like aida64 exactly how much you lose or gain by simply setting a few timings.

Set the trfc manually if you want to reduce latency, if you let the board "do it's thing" it will set a trfc of 350ns which is like 600odd trfc.

It's your computer not mine, I don't care how bad your latency or performance is.

Yeah, no, that's just not true. I've only set the tRFC and the others have different values.

On top of that, 1usmus has said that the sub tRFC values don't have to be changed manually.

 

[oobymach]

Huh?

The sub trfc are AUTO calculated from the main trfc setting. Your timings are all over the place. Those dimms can do cas 14, yet you have em set to 16. Then you set trfc to 288 which is cas14 range and run 1.4v. Like I don't get where you're going with these timings, much mismatched.

Da fuq u jus say? AUTO trfc is fine if you don't want to set it and gain some performance, mine AUTO sets to 600 odd trfc, I can cut that by more than half with a manual setting. Like I said in the end it's your computer not mine, I don't care how bad your latency or performance is.

Yeah, no, that's just not true. I've only set the tRFC and the others have different values.

On top of that, 1usmus has said that the sub tRFC values don't have to be changed manually.

View attachment 180735

Your trfc timings are wrong, but in the end it's your computer not mine.

 

[thesmokingman]

Da fuq u jus say? AUTO trfc is fine if you don't want to set it and gain some performance, mine AUTO sets to 600 odd trfc, I can cut that by more than half with a manual setting. Like I said in the end it's your computer not mine, I don't care how bad your latency or performance is.

Learn to read. The sub trfc timings are auto calculated from what you set the main trfc timing. Are you a tool or something? Why you attacking ppl trying to help you. Learn to post properly.

 

[oobymach]

Learn to read. The sub trfc timings are auto calculated from what you set the main trfc timing. Are you a tool or something? Why you attacking ppl trying to help you. Learn to post properly.

How am I attacking you? You're the one telling me to learn to read, calling me a tool for using a tool in the manner in which it was intended? Dafuq u jus say?

Idk, maybe I'm the only one who uses the tool to do the thing it was intended for.

 

[tabascosauz]

How am I attacking you? You're the one telling me to learn to read, calling me a tool for using a tool in the manner in which it was intended? Dafuq u jus say?

All anyone is trying to say is that before you bench AIDA/Geekbench3/membench and yield results that *show* a tangible, repeatable performance benefit to manual tRFC2/4, then all this that you believe is, well, what you believe. 0.5ns is within margin of error for AIDA and ~1s is margin of error for membench.

Literally no one brought up auto tRFC except you. I'm here making a point about tRFC being second only to tCL and you go off on a tangent about boards autoing tRFC to 600, like yeah, that's the point...

Here's what most people are going off of:



Stilt is Stilt and the timings system on Ryzen hasn't changed appreciably since Ryzen 2000.

I welcome your assertions if you give us some results that show an consistently appreciable benefit from manual.

 

[TheLostSwede]

Your trfc timings are wrong, but in the end it's your computer not mine.

Uhm... I don't have B-dies, so no, it's not wrong, but whatever...
Also, tRFC is set manually, tRFC 2 and 4 doesn't need to be set, as per 1usmus instructions. This is why it was moved to the Additional Calculator page, since it was on the primary settings page earlier.
This is from the version 1.3.0 release notes.

• TRFC2/4 now painted in gray color, that is, I do not recommend using them

 

[Glaceon]

I'm not exactly sure why any of you are manually setting tRFC2/4, that's literally not a "thing". Just set tRFC and let the board do its job.

OCD. I know, I still want to set it anyway.

I have yet to test it during games. But I did just wake up to 4 passes in memtestx86

 

[oobymach]

Uhm... I don't have B-dies, so no, it's not wrong, but whatever...
Also, tRFC is set manually, tRFC 2 and 4 doesn't need to be set, as per 1usmus instructions. This is why it was moved to the Additional Calculator page, since it was on the primary settings page earlier.
This is from the version 1.3.0 release notes.

You probably don't need to do this and it only really saves a few ms and adds a few hundred mb to your bandwidth so in terms of you being able to notice it you probably never will but I have ocd and like I said earlier my board left on auto for those sets it to much higher values so I messed with the calculator a bunch and found the fastest I could run it gave errors so I backed off to 160ns and gained stability.

The ram you have listed in your system specs is viper and it's not samsung b-die? If it's micron try a 220ns or if that won't boot try 250-260. The values listed in your screenshot show the middle value being highest which shouldn't be which is why I said it was wrong. Also nothing matters so no worries if you mess with it or not.

 

[Glaceon]

Okay well seems to me like it's reaching 50C. I'll keep stressing it like this and see if it spits out errors. I hope not. I'm not about to attach another fan to my RAM.

 

[TheLostSwede]

You probably don't need to do this and it only really saves a few ms and adds a few hundred mb to your bandwidth so in terms of you being able to notice it you probably never will but I have ocd and like I said earlier my board left on auto for those sets it to much higher values so I messed with the calculator a bunch and found the fastest I could run it gave errors so I backed off to 160ns and gained stability.

The ram you have listed in your system specs is viper and it's not samsung b-die? If it's micron try a 220ns or if that won't boot try 250-260. The values listed in your screenshot show the middle value being highest which shouldn't be which is why I said it was wrong. Also nothing matters so no worries if you mess with it or not.

View attachment 180745

Neither, it's Hynix CJR. All their sub 4000MHz kits are Hynix, all the kits 4000MHz and faster are Samsung.

Again, I guess you can't read, since as I have pointed out, 1usmus says that he doesn't recommend using the settings and this is why they were removed. In my case I see ZERO difference in setting them manually or leaving them on auto.

 

[tabascosauz]

OCD. I know, I still want to set it anyway.

I have yet to test it during games. But I did just wake up to 4 passes in memtestx86

MT86 really doesn't work fast enough for testing memory configs. If you're after efficient, TM5 anta777 does the work in 1.5hrs and you can even verify again and be done in a hair over 3 hours. If you're after overnight, HCI gets to 5000-10000% coverage overnight. If you want max heat and memory controller stress, some Linpack or Prime95 LargeFFT does the trick

If you are running 4 sticks, you can probably disregard most of the tRFC talk and leave it at the 300 mark if it allows you to run lower voltage. Having 4 DIMMs and a GPU in the system is no joke when it comes to thermals.

 

[Glaceon]

The difference between tRFC 324 and 288 for me is 0.3 ns. If this or TM5 spits out errors while the GPU and CPU are also stressed I might just put it back to 324 and voltage to 1.35v then forget about it.

Welp Prime 95 Large FTT gets one of the sticks in the 54C range and two others in 52C. No crashes yet though.

I think it's time to drop the voltage and up tRFC anyway.

 

[tabascosauz]

Manual tRFC2/4 vs. auto, 3 run average on each. No difference on my other two AM4 boards, auto won here. Seems a good reminder as any that like the relationship between RRDS/FAW, the BIOS doesn't have to listen to you. If the timings you put in don't make sense, the BIOS will always train its own values. It just won't tell you, and will continue to show you what you entered, but the performance will reflect that.



@Glaceon the 50C is just a general tip, above which temp sensitive ICs like B-die, Rev.E and Rev.B *can* lose stability. It doesn't mean you will, but if you were already pretty close to the edge, going above 50C could be the straw.

If you're already sitting at 1.35V there shouldn't be much risk of losing stability. You only really feel the link between tRFC and VDRAM if you're pushing below 150ns; in the opposite direction, might be hard to find benefit by loosening tRFC to try to shave off voltage.

 

[Glaceon]

Just tested both tRFC/2/4 manual (324-241-148) and tRFC/2/4 Auto (324-468-288). This is what I got. Top is manual, bottom is Auto:

Seems like I lost a little bit of Read/Write/Copy while latency remained the same.

If you're already sitting at 1.35V there shouldn't be much risk of losing stability. You only really feel the link between tRFC and VDRAM if you're pushing below 150ns; in the opposite direction, might be hard to find benefit by loosening tRFC to try to shave off voltage.

That 54C was while stressing everything in my PC at 1.4v on DRAM. I have 4 sticks of Trident Z Neo. I dropped the voltage back to 1.35v to be safe. The memory is stable, ran countless memory tests.

However temp concerns me, because it's B-die. It still reaches close to 50C on memory stress testing, and might surpass it with the GPU in action. I hope it's not gonna be an issue on 1.35v.

My case airflow is good, I have a NH-D15S fan in the middle circulating air, and three 140mm Pure Wings 2 PWM High-Speed fans as case fans, one as exhaust and two as intake, in a Pure Base 500DX.

Did a cold boot test too. After unplugging the PC completely and turning off the PSU for 20 minutes, and putting it back in (simulating a cold boot problem I had with tRFC 288 @ 1.35v while cleaning my PC), it booted instantly with little to no time on the DRAM light. So I assume that problem is fixed on tRFC 324 @ 1.35v.