Getting obsessed with B-Die, still don't have 2 sticks to my satisfaction

[GrimReaper85]

Board is MSI Z590 Gaming Carbon. I'm doing subtimings the last 3 days. It's so hard on dual rank. Single rank was boot = stable. But here I'm getting errors even after 1,5 hours. So now I'm doing one at a time. And I have to set them much higher than single rank, for example tWR before 6, now even 12 isn't stable.
There's not that many anyway, if I can get stable tRFC, tRRD, tRRD_L, tFAW, tWR, tRTP, tCWL, tWRRDSG and tWRRDDG, I'm more than half way there.
It still puzzles me how board can set these on auto and it's always stable. It's like they programmed in subtimings for dual rank and the worst possible modules that exist, as board doesn't detect which ICs you are running and how good they are. And it actually doesn't set them that bad, so far I only managed to lower tRFC from 771 to 308 and tFAW from 53 to 34.

My kit may be good, but what's this, how is this possible?

https://imgur.com/Auff3Xj

I couldn't even get single rank stable at more than 4400 CL16. And tCWL 15, I can't even get to boot odd tCWL. Kit is F4-4000C17D-32GTRGB. What's also strange is that it's 10850K and not 10900K with only 1.33/1.30V SA/IO.

 

[Woomack]

At least at lower clocks, single and dual rank Samsung works at not much different timings. Maybe you are simply reaching the stability limit.
Btw. set tCWL = tCL. It always works and barely ever causes any performance drops. The only rule is that tCWL can't be higher than tCL (motherboards or won't boot or will lower its value to tCL or less). Typically, when you leave it at auto then the motherboard will set it 2-4 lower, depends on frequency, etc. It also depends on some other settings but in general is not affecting performance so as I said, it's safe to leave it at the same as tCL.

This screenshot doesn't say if it's stable or not. I can show you screenshots at DDR4-5800+ in AIDA64 but they won't tell you anything.
10850K are generally the same quality as 10900K. Many world records were made on 10850K. It's a matter of luck but generally, these two chips are a bit better than anything lower.

If you like to play memory kits and tweak everything then I recommend buying a higher OC series motherboard. Maybe not now but in case of the next upgrade. I'm usually getting ASUS ROG higher series like Maximus or Crosshair but recently they're weirdly expensive so I moved to MSI as Unify is overclocking as good but costs significantly less. I only had weird problems with the Z490I Unify but X570 and Z590 are amazing for memory OC. Recently I passed DDR4-6400 on Z590I Unify - https://valid.x86.fr/quwxlt. I still have problems stabilizing anything past DDR4-5200 on this memory and any reasonable timings, but Microns work at DDR4-5200 CL18 so I'm not complaining.
ITX motherboards are always a good idea and on AMD, B550 seems to OC better than X570 (ITX or ATX, doesn't matter, B550 OC better). For example, I can make a 100-200MHz higher memory clock on Strix B550-I Gaming than on Crosshair X570 Impact while the Strix is 50% cheaper.

 

[GrimReaper85]

As I said, odd tCWL doesn't even boot (I have 4400 CL17), it's apparently a common problem. But it works at 16. Board actually sets it at 18 at 4000 CL16 and higher, so I guess it can be higher than tCL.

I thought I won't be doing this again as I've spent too much time playing with DDR4, whole 5 months. I only spent 1 month for DDR3 and even set all subtimings, but there's only like 20 anyway, so that is worth to set. But it's true I had 6 different kits and I tested even each stick separate. If I buy one good DDR5 kit in first place, so far it seems I will be done with this 2x16GB in 2 weeks max, so that is acceptable.
If I look back in txt file of first 3600 CL16 kit, there's a lot of weird stuff I can't explain even with so far gained knowledge. There was OS file corruption which I now know how to recognize and solve. Then at higher speeds there is figuring out your IMC capability. Then there was memory training issues, since then I just have Fast Boot - Slow Training enabled.
With this first kit I was at first at 4000 CL16 and 4266 CL17. Then I started setting subtimings on 4266 17-17-17-37-2T 1.43V, VCCSA 1.30V, VCCIO 1.25V and I got BSOD after 30min, so I forgot that and set 4000 CL16. Now I know for sure required voltages were 1.40V and SA/IO 1.25/1.05V, so I had more than enough and it was really weird. Then after some 2 weeks I had again major problems with stability. After one week I ran sfc/scannow and not much problems since then. After that I only needed to run Dism commands 3 times, but since then I already knew right away to recognize OS file corruption.
Final OC of that kit was actually 4266 16-16-16 and 4400 16-17-17 and would probably work even 4533 on better CPU like current 10900K.
Looking back, just leaving it at stock 3600 CL16 or settling on 4000 CL16 would be good enough. There were just too many difficulties to warrant anything higher than that.

But I just had to go through all this. I had DDR3 on only 1600 CL9 for 3 years, but I think only the memory went bad (same G.Skill, so far DDR4 G.Skill kits seem bulletproof, but that probably depends more on ICs themself) and I could have just bought another kit. Otherwise it was working 2133 9-11-10 for a few months.

 

[Brutal-Force]

GrimReaper,

I am curious. Of all of these speeds/timings posted, how many of them have passed at least one hour of Memtest, Karhu's, P94 or OCCT ram tests? I have a Gskill 4266 17-18-18-38-1T kit that is not XMP compatible or QVL compatible with my board.

I was able to run:

3800 15-15-15-36-1T for Geekbench 3
3800 13-13-13-30-1T for A64 memory and cache bench
4000 16-18-18-748-2T for A64 memory and cache bench

None of the above are memory test stable for longer than 20 min.

I ended up being 4 hours memory test stable at
3800 16-16-16-44-374-1T

I am not fortunate to have a motherboard that is very good with memory overclocking. Comparing with my kit, your numbers look pretty good. For the voltages you are running, 4000 at CL 16 is good as long as you can pass memory tests, if you can hit 4000 @CL15 and pass memory tests, then I think that is about as good as I would expect at stock voltages. I have hit 3800 @13-13-13-30-300-1T, but that is with 1.8V for benching and definitely only bench stable.

 

[GrimReaper85]

I'm always testing stability, P95 Large and Karhu on dual rank 2x16GB 4400 17-17-17 and 4533 18-18-18 are stable at least 8 hours on Auto subtimings.
I can't pass 4000 CL15 as CL15 doesn't even boot for me.
I will make some stability screenshots once I'm done with everything.

 

[EarthDog]

I am not fortunate to have a motherboard that is very good with memory overclocking.

Its the platform, bud (I think woomack said that?). A better mobo likely won't take you much further.

 

[Brutal-Force]

Its the platform, bud (I think woomack said that?). A better mobo likely won't take you much further.

Thanks. After a very lot of reading, and comparisons, for the platform I have and the kit I have, my speeds/timings are respectable. I can't complain.

 

[Woomack]

You still have a better CPU than the forum average. I remember some guys couldn't stabilize RAM at DDR4-4000. As I said, my 7900X can reach 4100 and nothing more (stable more like 4050 max as I remember) while 7800X can reach 4200. G.Skill was selling their highest memory kits for X299 at DDR4-4200 with info that not all CPUs will work at this clock.
Count that most motherboards can handle a 4200 memory clock but most IMC can't.

Try to work on the mesh clock instead of the RAM clock. If you had quad-channel then at ~3600-3800 CL14-14-14 with 3200MHz mesh clock, you would see around ~100GB/s in AIDA64 bandwidth and ~50ns latency.

 

[GrimReaper85]

Yesterday I made up what I lost previous 3 days with constant instability. 6 hours stability on already quite low subtimings, but some can still get lower. It was the voltage. Before I just tested 1.46V for 1 hour on primary timings and presumed 1.47V will be stable, which actually is. But for tightening subtimings I thought I could go lower with some timings if I give it 1.50V. But that made it unstable. So now I am at 1.47V.
Otherwise highest stable on dual rank I got was 4200 CL16 1.50V, but I only tested for 1 hour.
On single rank highest stable was 1.52-1.54V. I even tried some easy setting like 3600 CL14 which otherwise requires around 1.40V, and highest still stable was 1.59V.
I also remember one BSOD when I mistakingly gave it around 1.40V instead of required 1.30V. So I guess best stability is around lowest stable voltage.
But this problem with high voltage instability would probably go away with lower temperature. Single rank was at 44°C and dual rank is at 59°C.

Apart from those showed in ASRock timing configurator, I also lowered tXP to 1. tCCD and tCCD_L couldn't get any lower than default 4/8.

Once I'm done with secondaries and terciaries, I have to go through RTLs and IOs. Last time I was doing them it wasn't really that hard. Same as always, lower is better, except IO Compensation (the one that's 21 or higher). I actually rather start with RTL Init (the one around 60-70) as once I started with IO Comp and got to 24, but then RTL Init couldn't go as low as with IO Comp 21. I don't even know which one is more important, but since then this is the order I take.

One other thing I have to mention is the damn CLR CMOS. I still have 2 pins like 20 years ago. Not even 3 pins and a jumper like on ASUS boards 10 years ago, no, I had to go through box with old computer stuff to find a jumper. And not to mention the position of the jumper, just left of the F PANEL connectors. And there are other pins around it also. Usually it's not much of a problem, but only because I have slim fingers. But one time jumper slipped and I had to go looking for it at the bottom of the case. And yesterday it slipped back on top of F PANEL connectors so I had to go looking for something slimmer than my fingers. Usually after removing it, I put it on top of F PANEL first, so I can reposition my hand and pick it up again because some cables are in the way. Anyway, it's a hassle.
That's why I was thinking getting some better ASUS or ASRock board next time, MSI really disappointed me here.
So far 15 CLR CMOS required on this board, 6 only yesterday as board doesn't automatically reset BIOS if you go just 1 too low with some subs. On previous Z490 Tomahawk I only needed 2 CMOS resets.
Then I just enabled automatic CLR CMOS in BIOS. It actually works quite well. I think I was on BIOS V1.20 instead of current V1.00 when it booted 4 times for 2 minutes in any case, even if stable settings.

For 4600 I thought of trying 4600 19-21-21-42 1.45V, SA/IO 1.45/1.30V. If that doesn't work, I'll forget it. I know G.Skill has XMP 4600 18-22-22, but that one is single rank and it works straight 17 anyway. But it might be that for dual rank you have to go higher with tRCD/tRP.
EDIT: 4600 doesn't work regardless of timings.

 

[MaddMutt]

GrimReaper,

I am curious. Of all of these speeds/timings posted, how many of them have passed at least one hour of Memtest, Karhu's, P94 or OCCT ram tests? I have a Gskill 4266 17-18-18-38-1T kit that is not XMP compatible or QVL compatible with my board.

I was able to run:

3800 15-15-15-36-1T for Geekbench 3
3800 13-13-13-30-1T for A64 memory and cache bench
4000 16-18-18-748-2T for A64 memory and cache bench

None of the above are memory test stable for longer than 20 min.

I ended up being 4 hours memory test stable at
3800 16-16-16-44-374-1T

I am not fortunate to have a motherboard that is very good with memory overclocking. Comparing with my kit, your numbers look pretty good. For the voltages you are running, 4000 at CL 16 is good as long as you can pass memory tests, if you can hit 4000 @CL15 and pass memory tests, then I think that is about as good as I would expect at stock voltages. I have hit 3800 @13-13-13-30-300-1T, but that is with 1.8V for benching and definitely only bench stable.

View attachment 213603

Are you part of our benching team or have access to that area? Check out my posts in there - I also explain how to configure windows memory usage

 

[GrimReaper85]

Best I could do on 4400 CL17. tWTR was actually set manually, but there is no difference anyway. I tried tWTR_L 7 and motherboard sets the same 8 as on Auto.
Now I'm also running Karhu during my work hours as some don't like Prime95 Large for memory stability testing. I also don't like it spits out error as long as 10-11 hours for some tertiary timings, like tRDWR 14, though one time it was only 3min. And RTL 67 instead of 68. I'm quite sure there are better programs out there.
Github DDR4 OC guide says you shouldn't be running tREFI 65535 24/7, but 40000. I have to read more on this.



EDIT: Karhu only 23min stable, now I have to do binary search to find too tight timing(s).
EDIT2: Karhu nothing works at 4400 CL17, even just primary timings and additional +0.01V. I'm quite sure above settings are very stable, so I will first try 4200 CL16, then XMP and finally TestMem5.
EDIT3: I think it's temperature, lower frequency and lower temperature is more stable (4200 CL16 error after 55min) and 3600 is completely stable. Subtimings are fine, I used same (tREFI 35535 was a typo in BIOS) for 3600 and it is stable.
I doubt RAM will go to 58°C in real world, so I will just keep 4400 CL17.



EDIT4:
TestMem5 is also demanding:

 

[GrimReaper85]

What about 2x16GB 4000 16-16-16-36 1.40V? Voltage is really low so I think they're same for dual rank as 3600 15-15-15-35 1.35V is for single rank.
Should be really top bin and quite low price. I got 1.40-1.45V with my single rank kits and 1.42V with dual rank for 4000 16-16-16.
But I doubt there's really any other gains than -0.05V or so needed for 4000 CL16, at least on air cooling.
3600 14-15-15 are still more readily available. And 3800 14-16-16 and 4266 17-18-18 are nothing special with their 1.50V, I doubt they're any better than 3600.


I still want to be Karhu stable and so far I'm here:


So far I only know this errors right away, so upping one of these must have helped, probably tCWL:
tRFC - 280
tREFI - 65535
tCWL - 12
For more than 4 hours stability, I don't know, maybe up the voltage to 1.46V. Subtimings will be harder to pinpoint if some are still too low.

But this MSI Z590 Gaming Carbon is not that good:
BIOS V1.00 - PCIe speed was locked to 4 GB/s, so PCIe 3.0 x4.
BIOS V1.20 - triple boot
BIOS V1.30 - trouble with memory training, CL17 is fine, but on CL16 I always get different tWTR, tWTR_L, most terciaries and RTLs/IOLs.

 

[Woomack]

Z590I Unify is so far great for memory OC. Z490I Unify was meh as it had some weird clock walls on some specific IC. So far I went up to DDR4-6400 on a single Hynix stick (far from stable of course), DDR4-5200 stable with Hynix/Micron single rank, and DDR4-4800 with Hynix dual rank (seems stable but I wasn't testing it for longer). I wasn't testing Micron dual rank on this motherboard and my Samsung B dual rank runs at max DDR4-4266 so is nothing special. On the other hand, I have a new Team Group FPS 2x8GB DDR4-4000 CL16-18-18 1.45V kit and it runs at max DDR4-4800. It can't boot at DDR4-5000 but with some other Samsung B, I had the same in the past. I think that only TrodentZ Neo went up to DDR4-5100 at CL20 but other, single rank Samsung B couldn't pass DDR4-4800/4866.

I know you talk about the lower clock and I talk mainly about higher I just see no point to set something like DDR4-4000 using single rank kits and I see that on new Intel, the higher clock is showing some additional improvements so I generally push for DDR4-4800+ on everything I have in my hands. Dual rank seems interesting at DDR4-4400+ and not many kits can make that. New Hynix looks pretty good. As I said, I was able to set DDR4-4800 CL19-25-25 what passes 70GB/s in AIDA64 and about 54ns latency. After tweaking should look good.

 

[GrimReaper85]

I tried all BIOS-es. I can't use V1.00 and V1.10, PCIe is locked to x16 1.1 instead of 3.0.

Here is V1.20:


And V1.30:


I like subtimings on V1.20 more, they're always exactly what I type in BIOS. It's just that weird triple boot, but I can wait.
Someone said V1.30 is better because it trains lower RTLs. I still know very little about these RTLs and IOLs. I only read this:
https://hwbot.org/newsflash/3058_ad...ory_rtlio_on_maximus_viii_with_alexaros_guide

https://www.reddit.com/r/overclocking/comments/j3h9ot

Manually I never had much success, IO Comp 21 is highest it goes, and RTL Init and IOL Init anything lower than Auto is also not stable. I have no idea what tRTL and tIOL are for, but I know I can't lower them manually, they're dependant of what I set Init values to.

Other subtimings are no problem, there's mostly just 10 as terciares are already good on default, I only have to test tRDWR 15-15-15 instead of 15-16-16. I am finishing with 4133 16-16-16 1.49V, Karhu 12h+ stable.
Then I will try 4400 17-18-18, maybe that will be more stable than 17-17-17. Anything else is worse than 4133 16-16-16 anyway, so I won't bother.


I also bought what should be my final DDR4 kit, F4-4000C16D-32GTRGA. It should be best G.Skill dual rank bin, I would rank them like this:
4000 16-16-16 1.40V > 4266 17-18-18 1.50V = 3600 14-15-15 1.45V = 3800 14-16-16 1.50V > 4000 16-19-19 1.40V > 3200 14-14-14 1.35V = 3600 16-16-16 1.35V

These BIOS problems look really bad, but if I look at this G.Skill's kit QVL, there's mostly just MSI boards. And I read before that MSI is best for memory OC.

 

[Woomack]

RTL/IOL count mainly for competitive benchmarking. I don't really care about it for anything else. When you have a higher OC motherboard then it usually trains these settings a bit tighter than on cheaper "gaming" series but it's not a rule. It's also no guide that will exactly tell you how you should set it as depends on some factors it runs higher or lower and mismatched settings don't want to boot. It's usually +/- 1-3 but still counts. When you run it at auto then you can probably drop it significantly like 5-6. Especially at a higher frequency, it's usually more relaxed than it's required.

I was testing Team Group Dark Z FPS 2x8GB DDR4-4000 CL16-18-18 1.45V in the last days and up to DDR4-4800, it's overclocking like G.Skill Trident Z Neo DDR4-3600 CL14-15-15 1.45V. Somehow G.Skill could boot at DDR4-5000 on some motherboards and I couldn't do that on the TG kit so far.
I guess I will play some more with Silicon Power/ Hynix kit when I finish other tests. Can be Hynix but it's dual rank and runs at DDR4-4800 so for sure can make some good results, I just have to play more with sub-timings as I have no idea how it will act at such a high clock.
Besides that, I'm starting to sell additional DDR4 kits as soon will be DDR5 and it's appearing already on some mobile devices. I sold DDR4-3600 CL15-15-15 already.

 

[GrimReaper85]

All these top B-Die kits are most sensible only for competitive benchmarking anyway.

I think something like this for 190€ is more than sufficient for work/gaming:
Crucial Ballistix BL2K16G36C16U4R, 3600 16-18-18-38 1.35V - Micron Rev. E
G.Skill Trident Z F4-3600C17D-32GTZKW, 3600 17-19-19-39 1.35V - probabably some Hynix
In short, 3600 kits, 4000 are already too expensive.

 

[Woomack]

Ballistix 2x16GB can be E die in dual rank kits and B die in single rank kits. Micron B die is probably the only single rank IC used in 16GB modules in mass sales. 2x32GB Ballistix DDR4-3600 kits are for sure Micron B as I have two of these kits. They're also overclocking high for such capacity. I could stabilize all 4x32GB at DDR4-4133 CL18.
Ballistix MAX is Micron B, DDR4-5100 kit is marked as Micron N but I haven't seen any other kits with this IC.

G.Skill GTZN are almost only Hynix. Everything at CL18-22-22 is Hynix C, Samsung C or Micron E and usually low bins.

I see it like this:
If you want good Samsung then you get G.Skill.
If you want good Micron then you get Crucial.
If you want good Hynix then you get HyperX (or Hynix gaming series).
Everything else is random or depends on a specific PN.

... and another way:
If you want tight timings but not necessarily high frequency then you get Samsung B
If you want high frequency but still want to play with sub timings and tighten them some more, then you get Hynix D
If you want high frequency at somehow reasonable timings and don't want to play with sub timings then you get Micron B

If not for the IMC and various other limitations then Samsung B wouldn't be manufactured anymore. In 2018 Samsung stopped production and renewed it some months later because of the high demand. It's like back in the DDR1 times when Winbond BH5 production was stopped and back sometime later as UTT.

Btw, I'm using 32GB Micron B @ 5200 CL18 in my gaming PC but I'm thinking to change it to 64GB Micron B that will probably run at about DDR4-4400 CL18 and I set ~45GB RAMDisk for games.

 

[Zerileous]

I believe I still have a couple sticks of BH-5 laying around

 

[MaddMutt]

Ballistix 2x16GB can be E die in dual rank kits and B die in single rank kits. Micron B die is probably the only single rank IC used in 16GB modules in mass sales. 2x32GB Ballistix DDR4-3600 kits are for sure Micron B as I have two of these kits. They're also overclocking high for such capacity. I could stabilize all 4x32GB at DDR4-4133 CL18.
Ballistix MAX is Micron B, DDR4-5100 kit is marked as Micron N but I haven't seen any other kits with this IC.

G.Skill GTZN are almost only Hynix. Everything at CL18-22-22 is Hynix C, Samsung C or Micron E and usually low bins.

I see it like this:
If you want good Samsung then you get G.Skill.
If you want good Micron then you get Crucial.
If you want good Hynix then you get HyperX (or Hynix gaming series).
Everything else is random or depends on a specific PN.

... and another way:
If you want tight timings but not necessarily high frequency then you get Samsung B
If you want high frequency but still want to play with sub timings and tighten them some more, then you get Hynix D
If you want high frequency at somehow reasonable timings and don't want to play with sub timings then you get Micron B

If not for the IMC and various other limitations then Samsung B wouldn't be manufactured anymore. In 2018 Samsung stopped production and renewed it some months later because of the high demand. It's like back in the DDR1 times when Winbond BH5 production was stopped and back sometime later as UTT.

Btw, I'm using 32GB Micron B @ 5200 CL18 in my gaming PC but I'm thinking to change it to 64GB Micron B that will probably run at about DDR4-4400 CL18 and I set ~45GB RAMDisk for games.

WOW^ thanks for the great break down.. IÂ’m still using my I-Pad -> got to talk the wife into getting me a laptop for when my rig is down

 

[magellan]

@Brutal-Force
Is a i9-10980xe capable of quad-channel? Is it my imagination that you're running it in dual-channel?