• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Help me fine tune this memory

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

Voodoo Rufus

Powder Junkie Moderator
Joined
Sep 20, 2001
Alright, so I have these Gskill Ripjaws 2x16GB 3200C14 sticks in my main rig. I want to bust the 50GB/sec at <50ns mark. I can't get it to boot at 3733 or higher yet, so can I do it at 3600 and be happy?

3600-16-16-16-35-630.png
DDR3600C16.PNG
DDR3600C16_2.PNG
 

Oldschool_OC

Registered
Joined
Jan 12, 2015
Just a ballpark figure here, but I would guess @3600 CL18 in a z390 your looking at ~33 to 35 GB/s at ~26ns.
Just an estimate based on benchmarks I have seen.
Doesn't look like even CL 14 would get you above 35GB/s.
 

Woomack

Benching Team Leader
Joined
Jan 2, 2005
Alright, so I have these Gskill Ripjaws 2x16GB 3200C14 sticks in my main rig. I want to bust the 50GB/sec at <50ns mark. I can't get it to boot at 3733 or higher yet, so can I do it at 3600 and be happy?

Do you have memory OC profiles in your motherboard's BIOS? I don't remember what was in Gigabyte Z390 motherboards. If you have then probably there is something like Samsung tight settings. They usually work for 8GB and 16GB modules (matter of voltages and luck to IC).

In short, check something like this:
1. Check if you can boot with command rate 1N at DDR4-3600, can be even relaxed timings, just to check if it works.
2. Check if you can boot at DDR4-3600 14-14-14-28 or 14-15-15-32 at 1.40-1.50V (1.50-1.55V is still safe for longer). If it won't boot then can try twCL (the same as CL but for writes) at 12 or 14 (14 is generally safe).
3. If you make the above work then can move to sub timings. Here I won't help you much right now as I'm at work but there are many threads about tightening sub timings on Samsung B. As I said, 16GB modules are about the same as 8GB. If you have OC profiles in BIOS then they suggest how tight you can set sub timings.

On Intel you can also try 3800 CL14-14-14/15-15-15/15-16-16 or 4000 CL16-16-16 at 1.45-1.55V. It may require higher SA/IO voltages like 1.35/1.25V.
 

Zerileous

Senior Member
Joined
Jun 21, 2002
My 4x8GB 3200c14 Trident Z kit (on AMD) ran 3600 14 15 15 35 (never tried 32 for the tRAS), tRC 50 and tRFC 300 command rate 1T at 1.5v. To get 3733 C16 I had to go to 2T which cost more in performance compared to C18 with 1T. As Woomack said I would try to get 1T going first. I was using real world benchmarks and didn't really see a lot of difference between memory settings, but finally settled on the 3733 timings in my sig. This is what I get using AIDA (yes I know it's sad to use the trial version, but I almost never use the program)
aida.JPG
mem.JPG
 

Brutal-Force

Member
Joined
Feb 22, 2009
You could do 3600 @ Cas 16, but couldn't do 3700 @ Cas 17?

I usually make a chart. I have different combinations which show tighter timings in order. I try the combination of voltage, speed and cas with all other settings on auto initially. Then I tighten the other timings individually.

I would think with a 3733 18-18-18-38 kit, you should be able to either go up to 4000 @ 16 (if your memory controller can support it) or 3800/3733 @ 17-19-19-39.

Formula for the ratio is (2000*CAS)/Speed

excelchart.PNG
 

Woomack

Benching Team Leader
Joined
Jan 2, 2005
I'm not sure what this ratio is showing as nowadays you can't really count it in any way to show better or worse performance. There are too many variables and clock+CL are not saying much.
 
OP
Voodoo Rufus

Voodoo Rufus

Powder Junkie Moderator
Joined
Sep 20, 2001
So I goose the volts a bit and see if I can snug down the timings at 3600? I'll see what I can do.

Flashed to the latest modified bios for the board this morning. Still trying to get my settings back to where I like them. This will probably be a multi-day affair.
 

Brutal-Force

Member
Joined
Feb 22, 2009
I'm not sure what this ratio is showing as nowadays you can't really count it in any way to show better or worse performance. There are too many variables and clock+CL are not saying much.

In my limited experience, the ratio really only shows you some of the possible combinations. It does not necessarily indicate better performance. Reasoning: although higher speed and tighter timings might be achieved (can boot into windows), There are probably persistent errors which you really only see testing the memory. A benchmark isn't necessarily showing those errors (unless it crashes, like Geekbench). IMO, if you can boot into windows and run Karhu or Memtest for about an hour, then you will probably see good performance as well. To me, it also appears that speed>tighter timings, but that depends on your benchmark. A higher speed on Geekbench could be a lower score on Aida64 (depding whether you are testing reads/write or latency, and etc....)
 

Woomack

Benching Team Leader
Joined
Jan 2, 2005
To me, it also appears that speed>tighter timings, but that depends on your benchmark.

This is because higher frequency directly affects bandwidth and latency. Most timings give barely anything, some timings affect bandwidth and latency but more latency. This is why you should go for the high clock and next play with timings. Eventually, you will find the balance between clock and timings that will give you both, high bandwidth and low latency (usually it's somewhere at about 75% of the max clock).
On the way is usually standing the memory controller and its possible limits. On AMD also IF clock.
When you compare performance in the AIDA64, look at memory copy result and latency. Read and write isn't so important. When you hover a mouse cursor over the copy result then it actually tells you that.
 

Brutal-Force

Member
Joined
Feb 22, 2009
This is because higher frequency directly affects bandwidth and latency. Most timings give barely anything, some timings affect bandwidth and latency but more latency. This is why you should go for the high clock and next play with timings. Eventually, you will find the balance between clock and timings that will give you both, high bandwidth and low latency (usually it's somewhere at about 75% of the max clock).
On the way is usually standing the memory controller and its possible limits. On AMD also IF clock.
When you compare performance in the AIDA64, look at memory copy result and latency. Read and write isn't so important. When you hover a mouse cursor over the copy result then it actually tells you that.

Thank you for the added information. Whenever read stuff like this, it makes me try even harder. :)
 

MaddMutt

Member
Joined
Jan 12, 2015
I have a 9900kf with a EVGA z390 Dark/MSI z390 Ace and the best I can do is 4000 @ 14-14-28 1T. Are you looking for one time quick bench and then lowering back for a 24/7 setup?
It's all up to the IMC but if you want to try here is a guide for OCing the z390 Dark and B-Die memory. YOU HAVE BEEN WARNED THAT THIS MAY DAMAGE YOUR SYSTEM!!
^ The settings also worked on my MSI ACE board with Team Group DDR4-4133 (B-Die) -> These settings are for quick benching and not 24/7 use!!!

https://forums.evga.com/EVGA-Z390-DARK-and-BDie-based-low-latency-memory-overclocking-m2913749.aspx
 
OP
Voodoo Rufus

Voodoo Rufus

Powder Junkie Moderator
Joined
Sep 20, 2001
Nah, just looking for a nice 24/7 solid setup. Casual gaming is the highest workload it sees. But I like nice round numbers for benches and settings, hence the 50GB/sec target.

Or I could just get faster ram and let XMP hash it out with the MB.
 

Brutal-Force

Member
Joined
Feb 22, 2009
I have a 9900kf with a EVGA z390 Dark/MSI z390 Ace and the best I can do is 4000 @ 14-14-28 1T. Are you looking for one time quick bench and then lowering back for a 24/7 setup?
It's all up to the IMC but if you want to try here is a guide for OCing the z390 Dark and B-Die memory. YOU HAVE BEEN WARNED THAT THIS MAY DAMAGE YOUR SYSTEM!!
^ The settings also worked on my MSI ACE board with Team Group DDR4-4133 (B-Die) -> These settings are for quick benching and not 24/7 use!!!

https://forums.evga.com/EVGA-Z390-DARK-and-BDie-based-low-latency-memory-overclocking-m2913749.aspx

Thank you for adding this, too often stats/specs/accomplished are posted without context. It makes me feel inadequate when I see others accomplishments that make mine seem pathetic, only to see they have overvolted or run the risk of damaging components in the process.