XMP 14-15-15-35 3600Mhz B-Die won’t OC?!

[xg32]

3 hours ago, freeagent said:

That’s awesome man! Not sure why I thought you had an amd.

i'm about to build, just missing cpu and mobo, just waiting on that 5950x to drop to 800

 

22 minutes ago, Crestxo said:

That latency is sexy was that a benchmark OC or actually daily? 

that was the cl14 1.6v oc, the daily isn't much diff, cl14 3733 1.49v, latency goes to 38ns

[Guest]

If you are aiming for 4000mt/s with all 4 sticks, try 19-19-19-39 1.40v (Gear Down Mode Disabled + Power Down mode disabled.)

 

I wonder what PCB revision your sticks are. If they are A2, at higher speeds like 4000mt/s, using a 2T command rate timing may fair better for stability.

 

Most of all the other timings would be cutting edge competitive tweaking. This path takes weeks of learning, tweaking, testing and lots of trial and especially errors.

 

I suggest you make the minimal adjustments to start with. Worry about secondary and third timings later. 

[Crestxo]

12 hours ago, Mason Peterson said:

Yes, testing FCLK stability comes first. 

This is what I mean by SSE OCCT. Run it for an hour. 

  Reveal hidden contents

 

This is big brain time. DRAM Calculator is crap

Lucky for you, if you send more vDIMM, the timings will all scale lower (in contrast to Micron chips). That means tRCD, tRP, tRC, etc. etc. will go lower if more voltage. 

Relationship wise, not something super strict, just some general stability things in general. You can probably hit sub 150ns on tRFC ( timing / frequency * 2000 = XXXns), but it has no rule so test on your own. If you manually set tRFC, then don't even bother to touch tRFC2 and tRFC4, they won't be used.

tRRD_S two or three clicks lower than tRRD_L is general what I see (so think 4/6 or 6/8). 14-18 tFAW is what you're going to hit @ 4000. Since you're on AMD, tREFI is locked and so you can't touch it. 

Oh, and tRAS = tCL + tRCDRD + 2 is false information. The row doesn't "get closed early" if you go below that. tRAS below tRCDRD + tRTP, just you won't see any boost in reads. Still can see performance increases, go as you will (next lines as well). tRTP is going to always be beneficial lowering it until beyond a certain point because it will introduce heavy micro-instability (causing parity / ECC to hop in). 

tRC = tRAS + tRP is somewhat true. Going below that, you will almost never see performance uplifts. Does it mean it's worthless? No, lower timings overall will train faster and if you're doing three rounds of TM5 (rar file) of anta77 profile, it'll finish faster.

As you lower tCL, you should also lower tCWL (either == tCL or == tCL - 2. Depends on how low tCL you go). You lower it too much though, then tRDWD needs to increase, but that's actually fine because boost in writes! tRDWR burst gap = tCWL - tCL + tRDWR, so when you lower tCL without changing the other two, the burst gap will change.

That being said, you need to run an even tCL (14/16/18) with GDM enabled. It's just syncing the half clocks with how Ryzen does its memory controlling. GDM off is harder to stablize, and generally if you can do 4000CL18 with GDM enabled, try 4000CL16 not 4000CL17 GDM disabled. 

tRCD can be split into tRCDRD and tRCDWR. tRCDRD is what people typically associate tRCD with because tRCDWR is wacky. It doesn't linearily decrease as it should. I ran some Ballistix (Rev.E) @ 3733. tRCDWR = 17 was fine, = 15 was not, but 8 (this is the minimum setting in BIOS or in other words, this is the CPU registry limit)/10 were. Free performance as well. I can dig up more information in the JEDEC spec (since that's where the row access delays and how everything reads and closes actually comes from).

 

This is a whole fiery of information, so if you like a somewhat decent guide (but plenty of holes to poke), use Integral's (the notes on how to OC RAM is very much broad because it encompasses every type of RAM IC and speed bin). To look at your timings, use ZenTimings. When testing timings, I run TM5 for the full three cycles (will take ~3hrs). If I want to make sure I am 100% stable and can use it daily (this is why using multiple profiles is recommended: one for XMP, one for 24/7 stable OC, one for OC-in-progress), I'd run it for more than 6 cycles taking around 9-10hr overnight to complete.

 

There's tons of stuff to go on and on about, but it's not relevant right now. Loosen everything, work FCLK. Once done, come back and start working on primary timings. Then move to tRFC and secondary timings.

[5:55 AM]

Had some spare time earlier so I gave it another shot and followed your advice.

 

I decided to start with 3800MHz for now and try to get that stable since it's probably easier for a beginner like me, compared to 4000MHz.

First successful boot was with your suggested primary timings (16-17-17-37) as well as 1.05V for VDDG IOD and 1.00V for VDDG CCD - VDDP I left on AUTO for now.

Also as per your recommendation I set GDM to ENABLED and CommandRate to 1T - besides that I also put PowerDownMode to DISABLED.

As for DRAM voltage I got it to boot at 1.48V with SOC at 1.15V - RAM temperatures were between 40.2°C - 43.9°C.

It crashed after around 10 min (no BSOD just restarted (if that makes a difference)). After the reboot I tried to see what errors SSE OCCT would give me and it instantly gave me 2 WHEA errors and crashed immediately (no surprise there since it already crashed without any stress test before that).

 

Now after that it just wouldn't boot anymore with those settings - just boot cycling 2-3 times and then safe BIOS mode kicked in.

I tried to increase DRAM voltage to 1.5V with SOC at 1.15 - no boot.

After that I tried 1.5V DRAM with SOC at 1.175V and 1.2V - no success there either.

Sadly didn't have any time left after that to mess around with it a bit more.

I got a screenshot of ZenTimings during the successful boot though - do you see any timings out of place?

Currently reading the DRAM OC guide you suggested - hope to find an answer in there to this confusion

 

 

 

 

[Crestxo]

9 hours ago, ShrimpBrime said:

I wonder what PCB revision your sticks are. If they are A2, at higher speeds like 4000mt/s, using a 2T command rate timing may fair better for stability.

According to ThaiphoonBurner they are revision A1 - but heard a lot about Thaiphoon just guessing that / being wrong when it comes PCB revisions.

So not sure it it is accurate or not - also I don't know of any other way to find out whether its A1 or not.

[Guest]

Curtesy the blokes from hwbot forums.

 

[Crestxo]

9 hours ago, xg32 said:

i'm about to build, just missing cpu and mobo, just waiting on that 5950x to drop to 800

 

that was the cl14 1.6v oc, the daily isn't much diff, cl14 3733 1.49v, latency goes to 38ns

Is such low latency just an Intel thing or is it also achievable on AMD? Meanwhile me struggling to make 3800MHz stable

[Crestxo]

5 minutes ago, ShrimpBrime said:

Curtesy the blokes from hwbot forums.

 

So I'd have to remove my heatspreader to find out / verify?

[Guest]

Just now, Crestxo said:

So I'd have to remove my heatspreader to find out / verify?

If you look close enough near the center of the stick, you can usually make out the difference. 

I've come across the thread, you can read it here. https://community.hwbot.org/topic/187637-a1-or-a2-pcb-easy-recognition/

 

B-Die scales extremely well with voltage. The typical amount is 2x8gb.

The memory doesn't hold back the clocks, the CPU memory controller and Infinity Fabric does (on AMD chips)

 

 

 

 

 

 

[Mason Peterson]

58 minutes ago, Crestxo said:

Had some spare time earlier so I gave it another shot and followed your advice.

 

I decided to start with 3800MHz for now and try to get that stable since it's probably easier for a beginner like me, compared to 4000MHz.

First successful boot was with your suggested primary timings (16-17-17-37) as well as 1.05V for VDDG IOD and 1.00V for VDDG CCD - VDDP I left on AUTO for now.

Also as per your recommendation I set GDM to ENABLED and CommandRate to 1T - besides that I also put PowerDownMode to DISABLED.

As for DRAM voltage I got it to boot at 1.48V with SOC at 1.15V - RAM temperatures were between 40.2°C - 43.9°C.

It crashed after around 10 min (no BSOD just restarted (if that makes a difference)). After the reboot I tried to see what errors SSE OCCT would give me and it instantly gave me 2 WHEA errors and crashed immediately (no surprise there since it already crashed without any stress test before that).

 

Now after that it just wouldn't boot anymore with those settings - just boot cycling 2-3 times and then safe BIOS mode kicked in.

I tried to increase DRAM voltage to 1.5V with SOC at 1.15 - no boot.

After that I tried 1.5V DRAM with SOC at 1.175V and 1.2V - no success there either.

Sadly didn't have any time left after that to mess around with it a bit more.

I got a screenshot of ZenTimings during the successful boot though - do you see any timings out of place?

Currently reading the DRAM OC guide you suggested - hope to find an answer in there to this confusion

 

 

 

 

Give me a second to write a proper response to this.
 

54 minutes ago, Crestxo said:

According to ThaiphoonBurner they are revision A1 - but heard a lot about Thaiphoon just guessing that / being wrong when it comes PCB revisions.

So not sure it it is accurate or not - also I don't know of any other way to find out whether its A1 or not.

If Thaiphoon says A1, it's A2. Nothing really you should be worrying about. Only A2 on older Z170/Z270/select Z370 boards had problems.

44 minutes ago, Crestxo said:

So I'd have to remove my heatspreader to find out / verify?

Nope, and while I don't have explicit knowledge with Trident Z, it's been known that you can rip off the memory chips before separating the adhesives from the chips and heat spreader. 

48 minutes ago, Crestxo said:

Is such low latency just an Intel thing or is it also achievable on AMD? Meanwhile me struggling to make 3800MHz stable

It can be achievable on AMD. From what I recall, the best 5800X was 45ns, but not like I care. Aida64 memory latency means crap for in-game performance if you're talking cause-and-effect. It tests the system while it's in idle, and don't go loading a stress test while doing it because that's not what I mean by "idle". So beyond a certain point, why bother getting latency numbers down, look for performance increases in games / applications. You know having programs like Antivirus and RGB programs add nanoseconds of latency hence why if you boot into Safe Mode and try this, you can shave off many nanoseconds.

[Crestxo]

19 minutes ago, Mason Peterson said:

It can be achievable on AMD. From what I recall, the best 5800X was 45ns, but not like I care. Aida64 memory latency means crap for in-game performance if you're talking cause-and-effect. It tests the system while it's in idle, and don't go loading a stress test while doing it because that's not what I mean by "idle". So beyond a certain point, why bother getting latency numbers down, look for performance increases in games / applications. You know having programs like Antivirus and RGB programs add nanoseconds of latency hence why if you boot into Safe Mode and try this, you can shave off many nanoseconds.

My reason for overclocking is not to get more FPS - obviously don’t mind though. I just would like to get the maximum performance for what I payed for - if it takes me days/weeks then so be it.

I do have my anti virus and iCUE running of course, iCUE not even for the RGB - primarily for the fan control.

His latency just looked impressive because it’s almost 1/2 of mine and I am simply excited to see how far I can push my kit 

[Mason Peterson]

2 hours ago, Crestxo said:

My reason for overclocking is not to get more FPS - obviously don’t mind though. I just would like to get the maximum performance for what I payed for - if it takes me days/weeks then so be it.

I do have my anti virus and iCUE running of course, iCUE not even for the RGB - primarily for the fan control.

His latency just looked impressive because it’s almost 1/2 of mine and I am simply excited to see how far I can push my kit 

Give VDDG IOD 1080, CCD 1040, tRDRDSC = 2. Do that, try again.

If that doesn't work, set VDDP to 0.95V. If no workie, then 1V VDDP.

It looks like it is training fine for auto impedanices, but 36.9 Proc ODT is relatively low, 40 ohms is usually what it's @ 3800 for that same board. Try that with with the above changes. If that doesn't work, raise AddrCmdDrvStr to 24.

If none of those work, you actually should let VDDG CCD auto itself, VDDP auto, and VDDG IOD back to 1050

[xg32]

7 hours ago, Crestxo said:

Is such low latency just an Intel thing or is it also achievable on AMD? Meanwhile me struggling to make 3800MHz stable

just intel but doesnt mean anything, each platform is relative to its own stock/oc