8 PACK MEMORY RANGE GROWING: SAY HELLO TO 8 PACK RIPPED EDITION & 32GB KITS!!!

Associate
Joined
29 Jun 2009
Posts
549
VSOC 1.05
both VDDGs and VDDP at 0.95 could help

A few notes about Infinity fabric clocking in case anyone finds them useful:

1) You can test your infinity fabric independently, it doesn't have to be in sync mode. Example - set your memory to 3200mhz and raise fclk on its own.

2) The most important voltage for infinity clock is vDDG IOD. This must at all times be below vSOC as IOD voltage is derived from that - around 50mv is a good range to keep so that if you get any dips in vSOC due to load it doesn't cause a dip in vDDG IOD either. Example: at 1.15vSOC you would use 1.1vDDG IOD.

3) vDDG CCD has a hard limit and a stability range, if you raise it too high or if it gets too low compared to IOD you will black screen. 850mv to 950mv seems to be the sweet spot for most cpus. This affects stability when everything is overclocked - if your memory, cpu and infinity clock pass stress tests independently but not together raising this voltage a little can help.

4) vSOC is safe at 1.2v for daily clocks as long as you aren't riding temperature limits. Some cpus really like voltage, I have a 3600x that has been running 1.25v vsoc / 1.2v vDDG IOD / 1v vDDG CCD for 15 months with no ill effects. An indicator of how much voltage your own cpu is comfortable with can be found by how much vDDG CCD voltage it will take before black screening. This specific 3600x will take 1.12v before it wont post, considerably higher than all my other cpus.

5) For Ryzen 3000 your Infinity fabric limit is a hard limit. When you reach it your CPU wont post no matter how much more vDDG IOD voltage you add. If it posts you can make it stable with voltage, its very much on/off for Ryzen 3000. This behaviour could be changed with future AGESA updates, I believe at the moment it halts post if it detects errors whereas Ryzen 5000 does not.

6) For Ryzen 5000 your Infinity fabric limit is not a hard limit. You can go beyond where it is fully stable and this will show itself in increasing numbers of WHEA errors the further beyond stable you go. These can be observed in hwinfo and are also logged in windows event viewer. Eventually your cpu will not post because the errors are too high. AMD are hoping to increase the stable limit for Ryzen 5000 in future bios updates, however it seems they all generate WHEA errors beyond 1900mhz at the moment. You can still benchmark over 1900, but the more errors you get the more error correction you get so eventually you reach a point where you get lower benchmark results. I would not daily a cpu at IF speeds high enough to be generating WHEA errors.

7) You can start with vSOC, vIOD and vDDG high and tune downwards once you find your limit, its far quicker and easier than trying to tune voltage upwards and gain stability.

8) AMD default voltages are as follows: Under 2666mhz: 0.9vSOC, 0.8vDDG IOD, 0.7vDDG CCD | Up to 3200mhz: 1vSOC, 0.95vDDG IOD, 0.75vDDG CCD | Up to 3600mhz: 1.05vSOC, 1vDDG IOD, 0.8vDDG CCD | Over 3600mhz: 1.1vSOC, 1.05vDDG IOD, 0.85vDDG CCD. These are the voltages supplied by full Auto settings and changing only memory clocks while in 1:1 mode.


My own personal method for tuning IF is as follows:

Start at 1.25vSOC, 1.2vIOD and 950mv vCCD, memory at 3200mhz and IF at 1800mhz. Raise vCCD 25mv at a time to get an idea of the particular cpu's voltage tolerance. Most will take around 1v before they fail to post, the higher the cpu goes normally the more IOD you will need to get high IF clocks stable.
Return vCCD to 950mv, then start raising IF clock 33mhz at a time. On Ryzen 3000 if you reach 1900mhz IF stop, you wont be able to go higher anyway. If you black screen at post before 1900mhz then you have found your IF limit. Out of my 4 cpus, only one (my 3900xt) will not do 1900IF, that one does 1866IF and no more so most should achieve it.
Once you have your IF limit then start lowering both vSOC and vIOD 25mv together checking if you can post each time. Once you cant post then you have your minimum and can stability test within the range of voltage you have now found.


Finally, vSOC itself is only useful by itself as a stability aid when running multiple ranks of memory or high density memory. Quad rank per channel (4 dual ranked sticks in total) will nearly always benefit from vSOC voltage alone when testing for memory clocks, Dual rank per channel (4 single rank sticks or 2 dual rank sticks in total) is usually fine at whatever vIOD forces you to set vSOC to. With only 2 single rank sticks fitted you can usually undervolt vSOC if your vDDG IOD voltage allows.
 
Associate
Joined
18 Sep 2020
Posts
13
Thanks MrPils,
this thread and your answers are invaluable for someone who was used to just enabling XMP on an Intel Platform.
I just finished a Ryzen build this weekend and I am starting to dial in the settings now.
What would be a suitable tool to test the IF for stability with and for how long would you run it before further adjustments ?
 
Associate
Joined
29 Jun 2009
Posts
549
Thanks MrPils,
this thread and your answers are invaluable for someone who was used to just enabling XMP on an Intel Platform.
I just finished a Ryzen build this weekend and I am starting to dial in the settings now.
What would be a suitable tool to test the IF for stability with and for how long would you run it before further adjustments ?

Gaming is pretty good at testing for infinity fabric stability - any audio weirdness or usb weirdness is most of the time down to that. If you want something you can turn on and walk away from then Realbench is pretty good too, an hour or two of that will give you a close enough indication of your fabric stability that you can tweak the voltage over time and in day to day use for stability if required. Watch for WHEA errors either in your event log or by using hwinfo, most of the time these show up well before any crash and will get you very close to your stable IF voltage before you need to spend much time testing at all.

I would first give yourself an idea of where your fabric speed will end up when you get your build done by doing some gentle testing, don't try and find your final stable fabric voltage until you're at 1:1 with your memory. If you are manually clocking your CPU do that before you do any kind of thorough stability test on your memory or fabric, but if you're going to use PB2 or PBO then you can skip straight to fabric/memory. Once you have a mostly stable fabric clock (no whea errors, pass 20 mins or so of realbench) you can start raising memory speed towards it while keeping your memory timings loose enough that double your fabric speed should be well within the capabilities of your memory. Once you have reached 1:1 you can start tweaking memory timings - at this stage if you hit instability you want to add memory voltage first to see if that resolves the problem. You may find you need a small bump more fabric voltage along the way as your memory becomes faster, but mostly you'll be adding memory volts. As a general rule a test fail means more memory volts whereas whea errors mean more fabric voltage - as you close on in your final voltages and your best memory timings though this distinction can become more blurred.
 
Associate
Joined
29 Jun 2009
Posts
549
Pretty pointless testing 2x8 for comparison. Needs to be 4x8.
Any test of 4 x SR vs 2 x DR is pointless as it will only ever be testing how good the particular board used in the test is at clocking 4 dimms vs 2 dimms. The result will have nothing to do with the ranks, in a neutral environment 4 x SR is identical to 2 x DR. The IMC sees no difference between 4 x SR and 2 x DR as the slots are connected to the same channels, the only difference comes from the boards trace layout and training optimisation routines. The question is are dual ranks faster than single ranks when timings and speeds are optimised, not are dual ranks faster than other dual ranks.
 
Associate
Joined
28 Oct 2009
Posts
2,307
Location
Dark Side of the Moon
@MrPils

Again I need to thank you.

For a little while I'd given up on my 3900x hitting 1900 IF but using you method, I'm there and all voltages less than my super tight 3666 setup.

I will continue to work on lowering timings but needed a semi stable machine to work on tomorrow. I'll keep this post up to date on where I settle and if you are happy, can I ask further questions if required?

Thank you again Sir!
 
Associate
Joined
29 Jun 2009
Posts
549
B-Die Rank Comparison post #1 - G.skill F4-4000C16D-32GVK Results

Results are here for the 2 x 16Gb Dual Ranked setup. This is a retail dual channel kit imported from the US, the label says it was manufactured November 2020 so its brand new B-Die.

Test Setup and Methodology

Hardware:
Ryzen 5 5600x @ 4.5ghz
Asus Crosshair VIII Impact @ 1900mhz Infinity Fabric 1:1 settings
G.skill F4-4000C16D-32GVK @ 3800mhz 1.5v with fully tuned manual timings

I have limited vDDR to 1.5v and frequency to 3800mhz. Higher stable Infinity Fabric frequencies may be possible with future AGESA updates, in which case this test will be revisited if the frequency gain is significant. Whilst the kit can easily run higher than 3800mhz, unfortunately as with all other Ryzen 5000 cpu's I've seen so far this results in ever increasing WHEA errors as you move further above 1900mhz IF. This particular cpu is capable of booting to windows at 2133mhz IF but cannot run benchmarks and fails validation at this speed. Benchmark max is 2066mhz IF and requires no more voltage applied than the fully stable 1900mhz IF setting these tests were carried out at. The cpu used is a 5600x which has been manually set to 4500mhz fixed ratio. There is no PB2 or PBO active to muddle results, the only variable is to be memory timings, speed and ranks based on each individual kits abilities.

Results - 2 x 16GB Dual Ranked B-Die (1900mhz Infinity Fabric / 3800mhz / 1.5v / Full Manual Timings)

Click the spoiler below to see the test results using the Gskill_F4-4000C16D-32GVK at optimal timings for 3800mhz 1.5vDDR.

CPUz for verification of memory, CPU, motherboard and frequencies along with primary memory timings:

uc


Membench default and easy tests, full memory timings used. Note saved results are the best I achieved on a 3900XT @ 4.5ghz and Gigabyte X570i with 2 x 8GB SR Patriot 4400mhz Viper Steels so are not directly comparable.

uc


Next we move on to AIDA64, MaxxMem2 and Geekbench 3 and 4. MaxxMem2 in particular shows a massive gain from the unified cache and single ccx structure of the 5600x when compared to Ryzen 3000.

uc


uc


uc


uc


Now we move on to some more CPU intensive benchmarks - Y-Cruncher calculating Pi to 500million, 1Billion and 2.5billion decimals. This is all performed with the CPU locked at 4.5ghz.

500M:
uc


1B:
uc


2.5B:
uc


The next test is the HWBOT x265 benchmark in 4k mode:

uc


And now we can finish up with Timespy and Timespy Extreme CPU Only tests:

uc


uc

Early Performance Indications - 2 x 8 GB vs 2 x 16GB

I did a limited pre-test of the G.Skill F4-4000C15D-16GVK kit, it directly swapped with daily driver Patriot kit with no change in timings, but allowed a vDDR reduction from 1.51v to 1.43v. It was also capable of running tRCDRD 14 where the Patriot kit was limited to tRCDRD 15 if voltage was left at 1.51v. The Patriot kit is known high quality, so for the G.Skill 2 x 8GB kit to show improvements upon that is promising.

The main differences between the Single Ranked and Dual Ranked kits in terms of tuneable timings were subtimings. The Single Ranked kits were capable of running Command Rate 1T whereas the Dual Ranked kit required Gear Down Mode enabled. Additionally Tertiary timings go lower on the dual rank kits, specifically TwrwrSd and TrdrdSd (SD = Same Dimm Different Ranks) can be 1 in single ranked kits as there is no second rank to swap to whereas with the dual ranked kits these would not go lower than 6 and 4 respectively. Sc and Dd (Same Channel and Different Dimm) Twrwr and Trdrd timings were able to run at 1 for both kits as we are only dealing with 2 modules in total. Were we dealing with 4 sticks these timings may have been limited similarly to how we saw with the Sd timings in the dual rank kit. These tertiary timings can be quite impactful depending on the application being run.

Whats Next?

This post will be followed up in the coming days by testing the 2 x 8gb Single Rank G.Skill F4-4000C15D-16GVK kit. Hopefully I can get that done by the weekend, however I am not a reviewer nor am I paid to test. I am doing this during my own "free" time so don't be surprised if there are some delays. I've invested a lot of time into getting this far with the results since Saturday, it required a full rig rebuild including water cooling loop to bring up to Ryzen 5000 standard with a motherboard that was suitable for testing memory (the previous mentioned Gigabyte X570i is awful at recovering from post failure and has bugs with setting memory voltage from saved bios profiles, literally the stuff of nightmares) plus around 5 hours to get the memory tuned to its maximum stable settings at 3800mhz 1.5v. I have some serious family time to put in before I'll be permitted to run the bench suite on the 2 x 8gb kit. the good news though is I already know pretty much where they will end up in terms of timings as I'm much more familiar with Single Rank B-Die and know they are slightly more capable than my previous Patriot kit. This gives me a head start of several hours in terms of tuning.

Is it looking like Dual Ranked will be faster?

Overall there is not much difference in timings between the Dual Ranked sticks and the Single Ranked sticks. They're far closer to each other than I thought (given my previous experience with an older Corsair 3333mhz 4 x 16gb B-Die kit) so its going to be interesting to see exactly how big a delta that makes to the next set of test results. I have a feeling the Membench tests in the Ryzen Dram Calculator magnify the differences between the kits by reporting just over 10% in favour of the single rank kits based on my saved 3900XT results in the screenshot (though I do suspect some of this will be down to the Dual CCD structure of that chip doubling write bandwidth). Again the saved results in the Membench screenshot are not directly comparable due to the CPU and motherboard difference.

Its going to be closer than I expected it to be between 2 x 16gb Dual Ranked and 2 x 8gb Single Ranked that's for sure. Its a shame I haven't used more Dual Ranked B-Die kits made in the last year or so to know just how good this kit is. I have a sample size of 1 within this 4000C16 bin so I have no idea where these fall on the quality scale.
 
Last edited:
Associate
Joined
29 Jun 2009
Posts
549
@MrPils

Again I need to thank you.

For a little while I'd given up on my 3900x hitting 1900 IF but using you method, I'm there and all voltages less than my super tight 3666 setup.

I will continue to work on lowering timings but needed a semi stable machine to work on tomorrow. I'll keep this post up to date on where I settle and if you are happy, can I ask further questions if required?

Thank you again Sir!

Well done dude, glad it helped you get there. I'm always happy to answer any questions I can, just quote me to draw my attention and if I can help I will :)
 
Soldato
Joined
26 Sep 2017
Posts
6,185
Location
In the Masonic Temple
I concur with @MrPils findings as by coincidence we bought the exact same ram and I've been copying his settings, to learn, and have reached some ridiculous low timings. Achieved sub 55ns latency and can actually feel it in windows being snappier, I've also gained about 10% on average in games. And especially in red dead redemption 2 and timespy. In call of duty it was more like a 18% fps increase!

These are all in comparison to the same settings on my 3600 cl16 crucial ballistix running at xmp on a
5800x clocked at 4.8ghz all core with a 3080
 
Associate
Joined
29 Jun 2009
Posts
549
I concur with @MrPils findings as by coincidence we bought the exact same ram and I've been copying his settings, to learn, and have reached some ridiculous low timings. Achieved sub 55ns latency and can actually feel it in windows being snappier, I've also gained about 10% on average in games. And especially in red dead redemption 2 and timespy. In call of duty it was more like a 18% fps increase!

These are all in comparison to the same settings on my 3600 cl16 crucial ballistix running at xmp on a
5800x clocked at 4.8ghz all core with a 3080

Nice dude, glad you're finally happy with the memory in your rig. The Ballistix rep is a bit overblown right now, especially if you're running it at XMP. Ballistix is definitely cheap though, you can buy 4 x 16gb and have change for the price of the 2 x 16gb G.Skill 4000C16 kit. Can't argue with it as a value choice, but those auto subtimings for Micron really slay raw performance.
 
Man of Honour
Joined
24 Sep 2005
Posts
35,487
@MrPils thanks for the epic post and the detailed info. I must confess that due to a laziness or perhaps even a dumbness of my part, I’m finding it a little difficult to decipher.

Feel free to tell me to **** off and read your posts properly, but there any chance you could provide a super higher level executive summary? As in, (a) what you tested and (b) what the results showed / what was better?

I’m asked to give executive summaries at work all the time and it drives me bezerk so I feel like a toad for asking :o

Sorry / thanks either way!
 
Associate
Joined
29 Jun 2009
Posts
549
@MrPils thanks for the epic post and the detailed info. I must confess that due to a laziness or perhaps even a dumbness of my part, I’m finding it a little difficult to decipher.

Feel free to tell me to **** off and read your posts properly, but there any chance you could provide a super higher level executive summary? As in, (a) what you tested and (b) what the results showed / what was better?

I’m asked to give executive summaries at work all the time and it drives me bezerk so I feel like a toad for asking :o

Sorry / thanks either way!

At the moment that's the first part of a comparison, the second part will be a matching 2 x 8gb single ranked kit from G.Skill (F4-4000C15D-16VK). The purpose behind it is a proper follow-up on the Gamers Nexus video that seems to have lit a fire under everyone with a PC due to its claims of dual memory ranks per channel being better than single memory ranks per channel under Ryzen 5000.
I've split it into two posts mostly so that at least its more manageable for me than trying to do that whole damn thing at once. Also anyone running 2 x 16gb b-die can performance compare against my result results and timings. This also gives everyone an idea what a high bin kit of 2 x 16gb Dual Ranked B-Die can do and whether they feel a change from their current setup is required. I would encourage anyone not to make the jump based on my results until the Single Ranked results are complete and posted.
Part 2 will be exactly the same format except this time with the 2 x 8gb kit so then performance comparisons can be done. I'll tabulate the results for an easy compare as well so that no scrolling of benchmark screenshots is needed - they're only there as proof the tests were done.

I've edited the first part a little to hopefully make it a little clearer what's going on - feedback appreciated :)

[Edits to initial post complete]
 
Last edited:
Man of Honour
Joined
24 Sep 2005
Posts
35,487
@MrPils thanks so much for the additional commentary, very much appreciated. No wonder I was getting confused, I was (stupidly) looking for part 2 within that post :o

Looking forward to reading part 2 :)
 
Associate
Joined
28 Oct 2009
Posts
2,307
Location
Dark Side of the Moon
So been playing a little more as time has permitted. At the moment I'm at:

15,8,15,15,30 @ 1.4v RAM with semi tightened sub timings which result in AIDA:

Read: 59271
Write: 58979
Copy: 54445
Latency: 63.3

If I can figure out how to stick pictures in to post now, I will.

@MrPils any suggestions on moving forward? Trfc is currently 270 worth lowering?
 
Associate
Joined
29 Jun 2009
Posts
549
@MrPils thanks so much for the additional commentary, very much appreciated. No wonder I was getting confused, I was (stupidly) looking for part 2 within that post :o

Looking forward to reading part 2 :)
No problem at all, if you did it no doubt others have too. I think spending some time on the layout was worthwhile and I prefer the newly edited post in terms of readability too. It puts less emphasis on the results and breaks down the wall of text effect somewhat, so better all around. I always appreciate feedback, I'm bothering with this to help people gain a better understanding of how memory tuning works. If I can get my point across in a more easily understandable way to a reader due to constructive feedback then you have helped me as much as I you. :)

So been playing a little more as time has permitted. At the moment I'm at:

15,8,15,15,30 @ 1.4v RAM with semi tightened sub timings which result in AIDA:

Read: 59271
Write: 58979
Copy: 54445
Latency: 63.3

If I can figure out how to stick pictures in to post now, I will.

@MrPils any suggestions on moving forward? Trfc is currently 270 worth lowering?

Looking nice there, slow progress when learning is the best way to go about memory tuning. Over time you'll build up experience and then you can start taking shortcuts without opening yourself up to a brick wall you didn't see coming :)

I highly recommend Zentimings, its a great app as it displays every timing accessible from bios (and a few that aren't) in a single lightweight window. Its a very readable format and a useful aid so that comparisons can be made. Its now been updated to work with Ryzen 5000 as well, unfortunately I only discovered this after making my initial post so it wasn't included there. I'll make sure it gets included moving forward though as its really useful for at a glance timing comparisons.

Link to ZenTimings App

Regards posting pictures on this forum, I use my google drive. I'll spoiler that below so it doesn't bloat this post too much:

Step 1 - Log in to your google drive and navigate to the image you uploaded that you want to share
Step 2 - Right click the image and click on "Share"
Step 3 - In the lower pop up window click on "Change to anyone with the link"
Step 4 - In the lower window copy only the alphanumeric id string out of the link
Step 5 - Paste the alphanumeric from step 4 onto the end of the following URL: https://drive.google.com/uc?export=download&id=
Step 6 - On the OCUK forum click the small "picture" Icon above the text pane you are writing your post in - its between the emoticon icon and the movie icon.
Step 7 - In the popup window paste the link you created in step 5.

So In short:

Get shared link - example: https://drive.google.com/file/d/1zpbTltw0QOAESIiNXEUIxUKts8jj_sL7/view?usp=sharing

Remove id part of url string - example: 1zpbTltw0QOAESIiNXEUIxUKts8jj_sL7

Append to https://drive.google.com/uc?export=download&id= - example: https://drive.google.com/uc?export=download&id=1zpbTltw0QOAESIiNXEUIxUKts8jj_sL7

Insert created url into forum Image tags

If you post up a Zentimings screenshot I can help you calculate your timings a little better. tRFC should optimally be a multiple of tRC which itself is an additive of tRAS, so without knowing your other timings I can't really advise you properly :)
 
Associate
Joined
18 Sep 2020
Posts
13
So far I am having not much joy with this ram kit (I have the 8pack 2x16GB).
To be completely fair- I cannot pinpoint if its the ram or something else, just want to cover my bases at this point.
Don't want to derail this thread into troubleshooting my PC.
Occasionally the whole PC just reboots while not under load - Kernel Power Event ID 41 which is in most cases followed by a WHEA error (ID18, Processor core - Cache Hierarchy Error).
Is that what I could expect to happen with improper ram timings / voltages or should I investigate elsewhere first ?

Reboots seem to happen even if I just use the XMP profile on the sticks, rest of the Bios is in factory default settings.
The memory was checked with Memtest86 - no faults found. I also ran the kit in another PC (8700k) for a few days at XMP with no issues I could see so I don't expect a hardware fault.
Latest tests I ran were OCCT Stress test (XMP settings) and2h stresstest of RealBench (@XMP) without issues. So far I have not seen a reboot in a scenario where the system is used for stress testing/ benching / gaming - just "idle" regular windows use.
 
Associate
Joined
29 Jun 2009
Posts
549
@MrPils As per below. Work of note and do not know why but the volts are set in BIOS as:

SOC: 1.050
VDDP: Auto
VCCD: .950
VIOD: 1.025

Thanks :)

If you can see that your bios voltages for vSOC etc are as per what you set in hwinfo then I would ignore what zentimings says. My voltage readouts in zentimings don't work at all as Valorant's anti cheat blocks them...lol. It doesn't use the best of methods for obtaining them so I would not be shocked if they are bugged somehow. It reads memory timings perfectly though, so in so far as what we are using it for its still a great tool.

Regards improving timings, see below. You can ignore the voltage dependant main timings part if you have already minimised them for your chosen voltage (1.4v) and move straight on to the next section.

To anyone else reading the below - the main timings parts are for Samsung B-Die only. Other IC's behave very differently.

Voltage Dependant Main Timings:

tCL - this is directly linked to memory voltage. Higher voltage will get you lower. Pick a voltage you want to run up to 1.5v (B-Die safe) and test for your minimum tCL at that voltage.
tRDRD - this is directly linked to memory voltage. Higher voltage will get you lower. Pick a voltage you want to run up to 1.5v (B-Die safe) and test for your minimum tCL at that voltage.

Main Timings:

tRP should just work at tCL -2. A Quick test should suffice here.
tRAS on B-Die can go down to 22 - voltage affects this ability though so reduce slowly and carefully. Pay careful attention to post and boot times, you should "feel" when this becomes unstable and move back up if so.
tRC = tRAS + tRP so a reduction in tRAS and tRP can reduce tRC. You don't need to test this, calculation will suffice.
tRFC works best as a multiple of tRC. At 3800mhz your TRFC range will be 200 to 250 depending on voltage and stick quality. I recommend testing this with a bootable usb stick as it can cause corruption if its too low. Moving in such large chunks though you are unlikely to hit the corruption range when reducing it and more likely to encounter a no post.

You will see a good gain if you can tighten up your IMC to Dram bus latencies:

First look at tRRDS, tRRDL and tFAW. Test tRRDS 4 (as you are already stable at 6) for stability. Next set tFAW to 16 and test stability. Finally walk down tRRDL from 9 to 8, then to 6, then to 4. Quick test each step then perform a final thorough validation test when you reach your minimal settings of all three timings. More memory voltage or reducing memory speed will get these timings lower. Sometimes slower speed and tighter timings is faster overall.

Next look at tWTRS and tWTRL. Test tWTRS 4 (as you are already stable at 5), then walk down tWTRL from 14 to 12, then 10, then 8. Below 8 is unlikely to work, but certain configurations will allow 6. Quick test each step and do a final thorough validation test when both are minimised. More memory voltage or reducing memory speed will get these timings lower. Sometimes slower speed and tighter timings is faster overall.

After that you can look at reducing the channel latencies for your IMC:

Reduce tRDRDSCL and tWRWRSCL. Change them both together, try both at 3 and quick test then both at 2 then quick test. Usually with these if it boots its stable.
If you are only using two sticks usually both tRDRDDD and tWRWRDD can be 1 (the DD part means Different Dimms on the same channel), quick test here is enough to be sure its ok.
If you are using single sided memory usually both tRDRDSD and tWRWRSD can be reduced to 1 (SD means Same Dimm Different Rank), quick test here is enough to be sure its ok.

Finally subtimings:

tRDWR and tWRRD work as a pair. See if you can reduce tWRRD to 1, if so you should see a nice gain in calculation based tasks such as pi and prime. This can show in games depending on how the game engine is written. Altering the ratio of these two timings sacrifices latency for bandwidth as you reduce tRDWR and raise tWRRD so the "best" setting here is per user. I find 9/1 works best for most games.
tCWL can be up to -4 below tCL, gains are minimal after going past 2 below tCL though and hard to thoroughly stability test.
tWR can always be equal to tCWL, going lower offers minimal gains and is hard to stability test.
tRTP can always be equal to tWTRL, going lower offers minimal gains.


[Edit: typo - corrected "tWR can always be equal to tCWL". Apologies!]
 
Last edited:
Associate
Joined
29 Jun 2009
Posts
549
So far I am having not much joy with this ram kit (I have the 8pack 2x16GB).
To be completely fair- I cannot pinpoint if its the ram or something else, just want to cover my bases at this point.
Don't want to derail this thread into troubleshooting my PC.
Occasionally the whole PC just reboots while not under load - Kernel Power Event ID 41 which is in most cases followed by a WHEA error (ID18, Processor core - Cache Hierarchy Error).
Is that what I could expect to happen with improper ram timings / voltages or should I investigate elsewhere first ?

Reboots seem to happen even if I just use the XMP profile on the sticks, rest of the Bios is in factory default settings.
The memory was checked with Memtest86 - no faults found. I also ran the kit in another PC (8700k) for a few days at XMP with no issues I could see so I don't expect a hardware fault.
Latest tests I ran were OCCT Stress test (XMP settings) and2h stresstest of RealBench (@XMP) without issues. So far I have not seen a reboot in a scenario where the system is used for stress testing/ benching / gaming - just "idle" regular windows use.

That sounds like your power supply to be fair, idle restarts usually are. As you have access to another PC are you able to do a temporary PSU swap for troubleshooting? It may be worth updating your motherboard bios if you aren't on the latest, though I'm hesitant to recommend that unless your board supports hardware bios flashback in case you get a restart while updating. You can confirm if your memory is stable by running Testmem5 in windows using the anta777 extreme preset. If you pass that then your memory is stable, it doesn't excessively load any other part of your system either so its as close as you can get to a pure memory test.

I would make your own thread for your issue as it's almost certainly a hardware problem rather than related to anything you can tweak or tune.
 
Associate
Joined
18 Sep 2020
Posts
13
Thanks. Ill refer to general troubleshooting then.
(It isn't the PSU btw, I exchanged that already)
Hope to be back soon so I can start with the overclocking.
 
Back
Top Bottom