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

[GrimReaper85]

Guy on G.Skill forum with F4-4000C16D-16GVK just said it works with these voltages, so slightly worse than F4-3600C14D-16GTZNB:
4000 16-19-19 - 1.38V (+0.02V compared to mine, so my assumption that this is a better bin was not correct. These XMP voltages are really strange, some have only 0.02V headroom and some 0.13V)
4000 16-16-16 - 1.38V (+0.02V)
4000 15-15-15 - 1.46V (+0.01V)

I still have to wait for F4-4000C15D-16GVK to come, but I can already safely say that best buy DDR4 2x8GB kits are, both 160-170€:
F4-4000C16D-16GVK 4000 16-19-19-39 1.40V
F4-4000C15D-16GVK 4000 15-16-16-36 1.50V

Unless you're satisfied with 3200 CL14 or 3600 CL16. They can certainly do 3800 CL14, 4000 CL15 and 4200 CL16 with around 1.48V. Maybe also 3900 CL14, 4133 CL15 and 4266 CL16. But for 4200 CL15 and 4400 CL16 you need one of the kits mentioned on previous page.
Or if you're an extreme overclocker, then you need best of the best anyway.

 

[GrimReaper85]

I gave up on 4000 14-14-14.
I tried everything, VDIMM up to 1.65V. I also tried VCCSA 1.45V and VCCIO 1.30V to see if this is harder on IMC.
4000 14-15-15 needs 1.46V. But 14-14-14 is just not possible. It's semi stable 1.48-1.55V.
Like I said, best setting for 10600K 24/7 is 4200 CL15. Needs 1.54V, but I have to make use of this almost 250€ kit.

 

[Woomack]

It can be a matter of your motherboard. Not all like tighter timings above some point.

 

[GrimReaper85]

No problem then. To be really worth it for 24/7, it would have to be 4100 CL14 anyway. And that's hard on memory IC also.

 

[GrimReaper85]

Now after testing the not very good G.Skill Ripjaws V F4-4000C15D-16GVK 4000 15-16-16-36 1.50V I'm back to believing you just need the top memory ICs to do anything MHz/tRDC > 280. Maybe at 4000 14-14-14 motherboard also comes into play, but it's still more likely it's just memory. To know that I would need the best 2x 8GB RAM that's out there. But I don't think there's much better than the very good G.Skill Trident Z Neo F4-3600C14D-16GTZNB 3600 14-15-15-35 1.50V, apart from F4-3800C14D-16GTZN 3800 14-16-16-36 1.50V which should be exactly the same bin anyway.

I ordered Trident Z RGB F4-4400C16D-16GTZR 4400 16-19-19-39 1.50V for my final single rank B-Die. Easy to find in Austria, but very hard in Germany who also send to other countries. I ordered the last one, I hope get it. At least MHz/CL is the best there is at 275. But MHz/tRDC is still very bad at 231. 4000 15-16-16 is the best at 250, but it still doesn't guarantee it overclocks well.

This is what I got so far with F4-4000C15D-16GVK:
3200 12-12-12 - 1.47V
3500 13-13-13 - 1.50V
3733 14-14-14 - 1.51V
4000 14-15-15 - 1.50V
4000 15-15-15 - 1.50V
4000 15-16-16 - 1.47V
4266 16-16-16 - 1.52V

This is the absolute max for CL12, 13, 14, 15 and 16. It doesn't scale with voltage very good. There's a cap at MHz/tRCD = 270. So far it looks it's even slightly worse than Ripjaws 3600 CL16 or 3200 CL14, especially on CL14 where it doesn't even do 3800. And on CL13 and CL14 it's even worse than my very bad Patriot 4400 CL19.
G.Skill and their XMPs, not one kit gives you an idea what to expect. Though this kit at least has one of the best G.Skill XMPs, so best for those who don't have the time and will to OC.
Still, from my limited experience G.Skill has the best binning and Woomack also confirmed that. They're just missing top of the line, like that announced Colorful 4000 14-14-14, probably 1.55V.

As for dual rank I also did some digging today and these should be the best if I ever decide to try that before going to DDR5:
F4-3600C14D-32GVK
F4-3600C14D-32GTZN
F4-3600C14D-32GTZR
F4-3600C14D-32GTRS
F4-3600C14D-32GTRG

I don't believe 4000 16-19-19-39 1.40V is as good as 3600 14-15-15 1.45V, I don't like high tRCD and tRP. I think that should be clearer once I get my 4400 16-19-19-39. If even that one doesn't compare to 3600 14-15-15-35 then it should be clear which is best.

 

[GrimReaper85]

4400 16-19-19-39 is impossible to get. Now I will prepare myself for 32GB upgrade, I will change Noctua NH-D14 for NH-D15. I don't even know what Noctua was thinking with D14, you can barely use the second slot with 44mm DIMMs.

But again many choices:
2x16 GB:
3600 14-15-15 1.45V
3800 14-16-16 1.50V (should be same as above)
4000 16-19-19 1.40V
4266 17-18-18 1.50V (impossible to find)

4x8 GB:
3600 14-15-15 1.45V
3800 14-16-16 1.50V (should be same as above)
4000 15-16-16 1.50V
4266 17-18-18 1.50V (impossible to find)

Both same for OC or is 4x8 GB better?

EDIT: Forget it, 4x8 GB are 500€. Best buy are these two 4000 16-19-19 for around 320€:
F4-4000C16D-32GVK
F4-4000C16D-32GTZR

 

[Woomack]

8GB modules are easier for OC ... but motherboards like 2 modules more. In the end, on most motherboards, 2x16GB will OC better than 4x8GB. Only the top OC motherboards may show better results with 4 memory modules but so far I could OC 2x32GB higher than 4x8GB on all my 4 slot motherboards.

 

[GrimReaper85]

I actually got F4-3600C14D-32GTRG. One of the last few pieces.

One other thing is about best mem test. I was talking about Prime95 Large in regard of memory stability testing on Reddit and got downvoted so bad. That's why I even bought Karhu and from what I can tell Prime95 really is the fastest to finding errors. At least due to too low voltage on only primary timings changed.
I'll see if I can make use of Karhu when doing subtimings, but I won't bother with TM5 and HCI anymore.
https://ibb.co/HT3GQyx

Only one thing that I still need to find is the best setting. But Large already proved best on DDR3. I tried 448-4096 FFT which is suggested by some and I got worse results.

One other thing I found. I mistakingly had my RAM size limited to 6,5 GB in msconfig. At first I was furious as I was testing RAM for the last few weeks. But I actually found out it found memory errors sooner than with the whole 16 GB which actually makes sense as it had smaller size to check.
But 4000 14-15-15-32 I was testing on F4-4000C15D-16GVK was so far hardest to find stable voltage. Otherwise I don't think I have to limit my RAM size just to find errors sooner.

 

[Woomack]

I usually run AIDA64 memory+cache tests for 2-3h, and it's enough. When there is any instability then it usually crashes in ~1-10mins. Sometimes there are single errors at about 30-40mins but almost never past 1h.
Good luck searching for good answers on the web. Most people are clueless and only repost other peoples' not so bright thoughts. After multiple reposts, others take it as a rule.
Most memory manufacturers are using HCI or something similar. If they found it good enough then I see no reason why it wouldn't be. In the end, you can find many other stability tests that make a similar job.

 

[freeagent]

I have to agree with the name of the thread. I only own two pairs of B-Die and am very new to DDR4 but they have been the most flexible memory I have ever played with. From 3600 12-12-12 to 4400 16-16-16 these things have been a ton of fun to learn/play with.

 

[Woomack]

Micron B is more fun for me. At least my two kits run between DDR4-3800 CL14-17-15 and DDR4-5300 CL18-22-20 stable, and up to DDR4-5600+ for quick runs. These kits are also 2x16GB. I only wonder how they will act on new chipsets and if there will be anything worth higher clocks before DDR5 hits the market. Right now anything past DDR4-4800 isn't really showing expected performance because of too weak memory controllers and/or other limitations.

 

[GrimReaper85]

Most DDR4 OC guides are worthless. Most reviewers or those who write magical guides on the web have no idea about RAM or overclocking.
Btw. for some reason 9th gen processors had everything past 1.45V SA marked in BIOS as red and 1.65V marked as "can instantly kill the CPU". 10th gen processors have the red bar set as high as ~1.8V. Both generations have the same IMC. I have no idea how high voltage is still safe as there is no source that will tell you that. I haven't heard about anyone who degraded the CPU running at 1.45V SA but who knows, not all share everything on the web. I was running tests up to 1.8V+ and my CPUs didn't degrade but I'm not recommending that. On 10900K, 1.40-1.45V is enough up to DDR4-4800 (at least on my CPU). On 10500, regardless of SA voltage, my CPU couldn't pass ~4600 and max stable on MSI Z490I Unify was 4533.

Now we know the limit. On 1.45/1.30V I had around 30 hours under load. But 1.50V SA was immediate degradation right after booting into Win, I wanted to try 4533.
For 4000, before 1.16V, after 1.26V. 4266 shouldn't yet be a problem, before it needed 0.1V more than 4000, so I guess lowest 1.35V and 1.40V for good 24/7 stability. But I can forget 4400 now. It was out of the question for 24/7 anyway, but now I can't even bench it.

EDIT:
It's all fine now, works again 4000 SA 1.16V. I just had to set Memory Fast Boot (MSI Z490) to Slow Training. Don't know what difference between Slow Training and Disabled is, but I thought slow is better.
Didn't have any problems going down from 4400 SA 1.45V IO 1.30V though.

Others on Reddit say max is 1.50/1.35V, and dangerous 1.55/1.40V. I will let others fly so close to the sun, this CPU just isn't capable of 4533 on safe level.

 

[Woomack]

On most motherboards, slow training causes retrain timing each time you start the PC. Quick training is like it takes values from the last full training and disabled is almost the same, just skips all and uses values already set in BIOS. Instead of hidden, pre-trained settings, it uses default sub-timings from the timing table with recommended values which are usually slightly worse.

Most "others on Reddit" have no idea how it works and just repost whatever they saw on the web ... usually just thoughts of other people who didn't perform any real tests. This is how the web looks like nowadays, barely anyone performs any tests.
In short, whatever you see on the web, take as probably, not a rule as in most cases, it was tested on 1-2 CPUs, not representative higher amount to make any statistics. As I said, I was going up to 1.8V+ SA without any issues on more CPUs without any degradation. It's not a good idea but I'm also not saying that 1.8V is safe.

The maximum SA/IO depends on the CPU. Somehow 10th gen is officially stronger (not on the Intel side but motherboard manufacturers) but most those who are pushing hardware some more will still recommend keeping ~1.4-1.45V SA and 1.35V IO as max for 24/7 work. On the other hand, check how high motherboards are setting these voltages. They wouldn't let to set it above safe values at auto settings, and in most cases, ~1.40V SA is max, sometimes 1.45V.

 

[GrimReaper85]

Now on 10900K I got a little higher with F4-3600C14D-16GTZNB:
4600 17-17-17-36 1.53V, VCCSA 1.35V, VCCIO 1.30V
Not much of a gain compared to 10600K, but at least 4400 CL16 is now at comfortable 1.25/1.20V instead of high 1.45/1.30V.
4800 probably needs IO 1.35V, but I decided to stop at 1.40/1.30V, I don't want to push the CPU too hard.

But this RAM is amazing, from 3600 all the way to 4600 and it probably works all the way to 5000+ if going crazy with voltages. One would have never thought looking at XMP alone. Too bad it will go out anytime soon, replaced with 32GB F4-3600C14D-32GTRG.
Btw, I really hate the clips on NH-D15, they make regular RAM changing a real hassle. D14 was miles better, but unfortunately obsolete now, can only really use the rightmost RAM slot, they put RAM slots closer to CPU socket by half of RAM slot sometime after P67. Plus I got 8 degrees lower with D15. For DDR4 I will stick with this 32GB kit anyway.

 

[GrimReaper85]

Posts are gone, but I still have questions about DDR4 ICs and other things although I already have enough of everything.

1. You said Samsung works 4800+ on 10900K. But I was asking for 24/7 and dual rank. I just can't get 4600 stable. I got to 4400 17-17-17 SA 1.25V, IO 1.20V and 4533 18-18-18 SA 1.35V IO 1.30V. I know ICs can't do 4533 CL17 same as they can't do 4266 CL16, but I'm quite sure they would do 4600 CL18 with better IMC, if there even is any better. I tried 1.40/1.35V and instant freeze in Prime. Maybe SA needs more, but by then I'm out of 24/7 range anyway.
And 4800 is far from 4600, so my guess would be at least 1.55V SA and 1.40V IO.
I don't intend to go for any records, I just want to know if I did everything I could with reasonable voltages which for me are VDIMM 1.55V (I didn't get any stability past 1.52-1.54V anyway), VCCSA 1.40V and VCCIO 1.30V. I think I did everything right, it's just that your 4800 is getting me to question myself.

2. You talked about other ICs. But I googled and didn't find much about that Kingston with Hynix DJR. And Micron is probably on Crucial kits, god knows which models.
And they're supposed to be good for 5000+, but I'm not really limited by Samsung ICs even now, but by IMC. So I would need a platform like Ryzen 5000 or Intel 11th gen which support 1:2 gear ratio. And on lower frequencies they probably aren't as good as Samsung.
I'm guessing they're also not that easy to find. You can find Samsung in every supermarket and they're easy to spot. But how do you know which kits have Hynix DJR? What to look for? I know G.Skill 3600 16-19-19 is certainly Hynix, but I think CJR and not DJR. And there's also a question of price. And if they all overclock well, how are they binned.
I get your point of view from an overclockers perspective where you need to get as high as possible. I'm more interested in what makes most sense price/performance for some regular gamer. If I look back at the Samsung kits I bought, I would say:
3200 14-14-14 for AMD, as they don't work 4266+
3600 16-16-16 for Intel as they work up to 4400, maybe even 4600, I didn't have 10900K back then. Not as good timings as on 3600 14-15-15, but close, for example 4400 16-17-17 vs. 4400 16-16-16.

3600 14-15-15 is already too expensive, especially 16GB which is 230€. 32GB at 360€ is better value, but I would first need to see results from 3600 16-16-16 which goes for 290€. Probably also needs up to +1 higher timings, same as on single rank and already 3600 14-15-15 needs quite high timings, so for 32GB I think it's worth to pay more.
Does Hynix or Micron do anything that would be better value for money? This is basically my main question.

Some guy here said he chose G.Skill Neo 2x16GB 3600 14-15-15. Looking at Geizhals now I think he chose very well.

 

[Woomack]

I haven't seen a question about dual-rank, more like in general how I set it so high. Anyway, now some posts are gone so we can continue.
1. My current dual-rank Samsung didn't like Z490 and now I see is the same on Z590. It hits a wall at 4133-4266, regardless of timings and voltages. It simply doesn't boot any higher. Typically, dual-rank Samsung doesn't even work at DDR4-4400.
2. You won't find much about other IC because people are not testing them. They just google for "the best RAM" or whatever and later find forum threads where people glorify Samsung B. Later in 99% of posts, you have only info about Samsung and that other ICs are bad so are not worth mentioning.

Hynix DJR = pretty much all new higher frequency kits, check top motherboard QVL and for example, MSI or Gigabyte add used IC, most of the 4400+ list is covered by Hynix DJR nowadays. Hynix is used by all well-known brands. Samsung is only in some "top" and overpriced kits. Everything less expensive is based on Hynix or Micron.
Micron E = usually 8GB single rank or 16GB dual rank. However, barely anyone is selling dual-rank modules on this IC. Micron B = 16GB single rank or 32GB dual rank. This is probably the only IC in 16GB single rank modules, also Crucial is selling them in DDR4-4400 Ballistix MAX kits. This IC made some world records and I pushed it to DDR4-5656 on Ryzen 4000.

I'm not saying that Hynix or Micron is better than Samsung. They're different and better or worse for some specific setups.
For example:
- Good Micron B IC runs at 5000-5400 CL18-24-24 1.65V, but good is available only in Ballistix MAX which is very expensive. All other brands which are using this IC have much worse bins. You can't really tighten the timings on this IC and if you do then not so much so it will be worth mentioning. At least in my tests, it's acting better on AMD than Intel.
- Good Hynix D runs at 5000-5600 but requires high voltages and more relaxed timings like 20-26-26 -> 24-32-32. Can play some more with sub-timings than on Micron and it runs better on Intel than on AMD.
- Samsung won't run at so high frequencies or even tight timings at DDR4-5000+ but is scaling much better with timings up to ~DDR4-4800 (dual rank till about 4400). Highly depends on the motherboard and IMC quality.

You can see what IC can be used in each memory kit by looking at the XMP profile and memory price.

You don't need 11th gen Intel to run at a higher memory clock. 10th gen handles anything up to 5000+. It's a matter of IMC quality, motherboard and voltages. Some CPUs can make 5000 stable, some won't pass 4600. However, I haven't heard about higher i7 or i9 that couldn't pass 4800. Problems are mainly with i5 and locked CPUs which are usually working up to 4600.

 

[GrimReaper85]

I actually presumed dual rank might be harder for IMC, but not as much as it actually is. It could also be the ICs, but with SA/IO 1.35/1.30V it didn't come to desktop and with 1.40/1.35V it was stable at least on desktop, so that makes me think it's the IMC. Good/bad I don't even know, I found much worse around the net.
First thing I need to do today is test the IMC on 4200, 4400 and 4533 to see how much more voltage than single rank it really needs. Then I can probably make an estimate of how much 4600 would need, unless there's a wall or some huge gap.
For single rank I also didn't get 4800 stable, but I only tried 1.40/1.30V.

Also one thing else I noticed on MSI Z590 Gaming Carbon, double boot when changing RAM speed. I tried all BIOS-es and there is double boot on V1.00 and V1.10 and even tripple boot on V1.20. No such problem on Z490 Tomahawk.

 

[freeagent]

I mainly run dual rank, and its tough. A lot is IMC, but I think board plays a part too, but I cant say for certain because I have very limited experience and time running. With this setup, when I wind up a pair of single ranks they can fly, and performance is excellent. But it just seems to be a little better paired up with another. With 2 pairs I can only push to 1900, with 1 pair I can push to 4400, maybe 4600 if I wasn't so shy with voltage.

 

[GrimReaper85]

Tested my 10900K IMC now on 2x16GB:
4200 16-16-16 - 1.11/1.16V
4266 17-17-17 - 1.15/1.04V
4400 17-17-17 - 1.20/1.14V
4533 18-18-18 - 1.36/1.21V
4600 19-19-19 - 1.40/1.30V or 1.45/1.35V - can't even get to Win, by now both IMC and DIMMs are probably holding me back, if not even mobo
I put +0.05V for 24/7 stability. It doesn't seem they are fixed values like for VDIMM, it varies day by day, especially VCCIO.

On AMD X570 + 5600X + 4x 8GB I just can't get more than 3600 14-16-14-14-32 to boot. Github guide says I have to set ProcODT. Now it's at 36.9 Ohm.
I also tried UCLK divider /2, so I think it's not the IMC. Single rank it goes up to 4000.

Anyway, I'm satisfied, both 4400 CL17 and 3600 CL14 should perform similarly.

 

[Woomack]

I don't remember what intel motherboard do you have but on MSI Z490I Unify, I couldn't pass DDR4-4533 with some Samsung memory kits. Like a hard wall at 4533 and not an MHz more. I have no idea why. On the other hand, your 2x16GB kit overclocks much above the average and I would stick with 4400-4533 and work on sub-timings if you want a bit better results.
One more thing. Many Samsung IC at higher clocks run more like CL18-20-20 or 19-21-21. You may try that but I guess that with 2x16GB you won't make much more anyway. At least it was working this way at DDR4-4700->5000 on some 2x8GB kits.

On AMD, at least in my experience, it's harder to OC 4 memory modules than 2. I had better results with 2x16GB or 2x32GB kits than 4x8GB. Maybe it's also a matter of motherboard and some other things.