[GUIDE] Skylake Memory Timings on Asus Motherboards !

[xxbassplayerxx]

I just beat a 6700K with HT disabled running 2 cores 2 threads overclocked on liquid nitrogen to 6159.6MHz with a i3-6100 with HT enabled 2 cores 4 threads overclocked to 4,801.16MHz on air by 31.4 Marks. It certainly seems like that invalidates a bit of your blanket statements.

 

I don't know as if it runs better with cores turned off versus on it's plausible though. You'll have less throttling to worry about that could skew results up to a point though that's true with any benchmark that's not overly multi-threaded as well.

 

http://hwbot.org/submission/3202227_invasmani_maxxmem_ddr4_sdram_3925.9_marks/

 

I don't think that does anything other than validate what I'm saying. You ran memory and CPU speeds that are massively lower than the existing first place and somehow beat it. The benchmark is fucked.

 

Your 3110 12-13-13 isn't somehow faster than l0ud's 3900 11-16-16. As more evidence that no one runs this benchmark, people are doing 4080 12-11-11 with B-Die but no one has submitted a run to beat l0ud's, even though the frequencies are possible.

[Guest invasmani]

It's a memory benchmark not a CPU benchmark the impact on the latter should be marginal at best. Also it's well known Skylake's IMC only scales so with pure frequency and in fact the sweet spot is right around 3200MHz past that point roughly the impact lessens quite heavily much like high frequency CPU OC and vcore.

 

Yes sure CPU speeds and memory frequency was lower, but it was also running 64GB in quad channel with tight timings rather than 8GB of dual channel. I don't know the impact on ram density nor quad channel in said benchmark, but I do know timings make a substantial difference.

 

For the record I beat him in the following

 

Benched in time 7sec 217ms 704us

Displayed in time 11ms 909ms

Reached memory score 32.80GB/sec

Latency score 27.3 ns

 

While he beat me in the following

 

Memory-Copy 49993MByte/sec

Memory-Read 39004 MByte/sec

Memory-Write 45362 MByte/sec

 

So I beat him in 4 categories he beat me in 3 is it much wonder I got a higher overall score particularly when my benched/displayed, latency and reached memory scores were all better?

 

So he beat me in sequential burst rate speeds which are the most CPU driven bottle necked portions of the entire benchmark test. Those are the scores that all tank the hardest when you lower CPU speed, but maintain the same memory frequency and timings. The fact is timings matter as much as frequency. Quad channel is also better than dual channel at the same timings which I'm pretty certain only became more true with DDR4. I'm sorry that you have a problem with the benchmark bottom line though I beat his result and that's what matters. I even saved the benchmark result file as well. Benchmark popularity aside I deserve my due credit it's a hell of a accomplishment and I'm proud of my results it was no easy task with the bar set that high even if it can be raised because the bar can always be raised that's the whole point.

[zeropluszero]

Its ok guys, I reported and scores have been blocked.

[Guest invasmani]

Yeah for what specifically I beat some someone in a memory benchmark. Please come to terms with and acknowledge the fact it is a memory benchmark. It's not as if I don't have proof and I submitted a screenshot same as everyone else and I even have the benchmark file as well on my computer's HD.

 

You have to realize my reached memory score was 8.8GB/sec higher with 7.2 ns latency less on top of that it benched over a full second quicker and displayed time was even faster yet oh, but I guess none of that matters because bug can't possibly be valid since it's a CPU benchmark...oh wait!

 

Useless benchmark or not I got the best results yet with it and that's a fact. Shame you don't have a valid reason for blocking the best you've got is "it's impossible" as that's the best you've got to come up with now if only that answer worked for everything.

[zeropluszero]

Thanks!

[Guest invasmani]

Here's a bit math for you that helps explain partly why my score was better despite a far lower CPU clock speed and that's on top of 7.2ns better latency.

 

6,159.6MHz/31x (at 100MHz) = 198.69677419354838709677419354839 (DRAM reference clock frequency per CPU MHz higher is better)

6,159.6MHz/34.5ns = 178.53913043478260869565217391304 (DRAM latency per CPU MHz lower is better)

6,159.6MHz/3908 = 1.5761668372569089048106448311157 (DRAM frequency per CPU MHz higher is better)

vs

4,801.16/24x (at 133MHz) = 200.04833333333333333333333333333 (DRAM reference clock frequency per CPU MHz higher is better)

4,801.16/27.3ns = 175.86666666666666666666666666667 (DRAM latency per CPU MHz lower is better)

4,801.16/3111.4 = 1.5430867133766150286044867262326 (DRAM frequency per CPU MHz higher is better)

 

1.5761668372569089048106448311157

-1.5430867133766150286044867262326

= 0.033080123880293876206158104883 DRAM frequency advantage overall

 

178.53913043478260869565217391304

-175.86666666666666666666666666667

= 2.6724637681159420289855072464 DRAM latency disadvantage overall

latency advantage

 

So now ask yourself what would you rather have 2.672 higher latency or 0.033 higher DRAM frequency "inconceivable" do the math it checks out just fine!

Edited May 3, 2016 by invasmani

[zeropluszero]

tl;dr

[Random]

Invas, that statement is for asus boards, they use T-topology so THEIR 4dimm boards are messy.

 

Asrock is completely different.

And if by some miracle you beat louds score, he probably would post a backup and beat it.

[Guest invasmani]

Invas, that statement is for asus boards, they use T-topology so THEIR 4dimm boards are messy.

 

Asrock is completely different.

And if by some miracle you beat louds score, he probably would post a backup and beat it.

It was a good generalized DDR4 reference guide regardless was just posting as a reference point with my hardware what worked and didn't from it.

 

From everything I've read they should just retire that particular MaxMEM benchmark and score results. I have it on good authority no one really uses it anymore and only the read performance matters so they should strictly use this one instead.

http://hwbot.org/benchmark/maxxmem_read_bandwidth/

[Guest invasmani]

Lol. maxxmem might be the worlds worst benchmark.

Gotta be the best at something buddy why not that!

[Random]

I was specifically referencing someone saying 2t is necessary for over xxxx i got on wrong page

[Schmuckley]

Man I came here for info on stabilizing b-die and all I see is dog-poo.

 

Yes maxxmem is a buggy bench.Be vewy vewy quiet lest it gets cut.

Better benchmarks have already.

[GtiJason]

Man I came here for info on stabilizing b-die and all I see is dog-poo.

 

Yes maxxmem is a buggy bench.Be vewy vewy quiet lest it gets cut.

Better benchmarks have already.

 

Really, we've known this for at least 5 years

 

http://forum.hwbot.org/showthread.php?t=22522

[Guest invasmani]

Really, we've known this for at least 5 years

 

http://forum.hwbot.org/showthread.php?t=22522

Two bugs to be precise moving the program window during the benchmark to a different spot on your desktop and changing thread priority or reopening the program prior to benchmark. No one cares if it gets cut it's buggy it's useless and they only care about the write score anyway.

[GtiJason]

Two bugs to be precise moving the program window during the benchmark to a different spot on your desktop and changing thread priority or reopening the program prior to benchmark. No one cares if it gets cut it's buggy it's useless and they only care about the write score anyway.

 

Yep, I tried your tweak the day you posted it. Too bad MaxxMem so buggy or you could have road into the sunset as king of MM !

[amirhosein]

maybe is late by th way taht is awesome Topic alex !

tnx bro

[speed.fastest]

This board is flying! Without your guide maybe i can't take it this far, thank you again :ws:

[Alex@ro]

No worries,have fun !

[Alex@ro]

Updated first posts with B-die info and experiences.

 

***

Since this guide was originally posted before Samsung B-die made an entrance in them market,I decided to add my experiences with it later on the path.

B-die a.k.a K4A8G085WB started manufacturing middle of 2015 with the first IC carrying week 524 code on Samsung OEM and 528 on 3-rd party vendors(according to my own research) .

It shares a lot with previous E-die in voltage/mhz/tcl scaling while also allowing ridiculous low TRCD-TRP limits which simply makes it the best choice in overclockers arsenal .

At the moment 70% or more of the market B-die seems to be manufactured at the same factory with the same 10-layer pcb as you can see in following picture , Teamgroup , Zadak511 and G.skill sharing same PCB together with Galax and possible others too.

 

 

As good as it is for benching ,this Particular IC has it’s flaws also .Those seem to be :

-Voltage tolerance

-copy-wazza issue

Regarding voltage tolerance B-die seems to be less permissive than E-die,with a lot of sticks getting issues at 1.95V and over .

Considering B-die voltage scaling is like 0.1V per each 133 mhz at TCL12(e.g DDR4-3733 12 12 12 required 1.67,DDR4-3866 requires 1.77,DDR4-4000 requires 1.87 on same stick) it is clear that the goal is to find sticks that are decently in voltage requirement at DDR4-4000 (under 1.95V) but tolerating 2.05 .Please note that voltage tolerance does not improve with cold or different cpu or motherboard.

Copy-wazza issue was discovered and currently seems to be the flaw of many sticks . The flaw it’s simple, module will run every benchmark at a given frequency and timings except S-pi 32M with copy-wazza tweak which required 200-300 mhz less for benchmarking. Based again on my own research it seems that the flaw is related to B-die production weeks and my testing showed that:

-sticks 534-540 including 540 will pass most of the times wazza ;

-543 has a 50% chance of passing wazza;

-545 549 552 will fail most of the times ;

-601-616 has a decent chance of passing wazza;

-619 will pass most of the times wazza (80%+ in my testing)

4-dimm motherboards will have a hard time running B-die over 3866 and if your goal is to reach DDR4-4000 12 12 12 on these motherboards ,serious binning of the motherboard is required.

Therefore almost all of my testing was performed on Maximus VIII Impact which can handle way better the 8GB sticks.

Similar to E-die , Samsung B-die allows very tight tertiary and secondary timings,compared to previous generation trfc can actually go very low like 180 at 4000+ without however any significant performance improvements in SuperPi. A value of 240 should be good match for stability and performance as some sticks can have problems regarding 200 or lower at 4000+.

The following combo should not limit overclocking or performance :

 

Please note in rare cases some sticks may not like Write Recovery time and TRTP,therefore don’t forget to try relaxing this as last resort.

An important note also is that tertiary TWRWR_dr and _dd actually benefit from values like 4-4 or 5-5 on Asus Maximus motherboards compared to rest of the IC’s tested.

Maximus VIII Impact can handle very tight RTL at 4000+,typical values are 49-50-6-6,50-50-7-6,50-51-7-7,51-51-8-7.

These advanced SLOPES settings may give you a little boost in improving stability and MHZ:

For my particular IMC/Motherboard combination the outcome at 4133 12-11-11 timings was the following:

-ALL AUTO loop 14-15 would crash;

-ALL SET 15-1 Initial round failed;

-ALL FALLING SLOPES 15-1 ,rest auto would fail to load windows;

-ALL RISING SLOPES 15-1 ,rest auto would pass loop after loop with wazza.

Also note that increasing slopes value will give VCCIO a better scaling.

 

DEBUGGING B-DIE

So you got yourself one or more B-die sets for benching and you want to check the potential on your Maximus VIII Impact. You might want to check on the following issues.

1. Eternal boot of the spotless B-die

Some sticks exhibit this issue even at DDR4-3733 12-12-12 ,they will need over 1 minute to start OS loading.However most of the sticks would manifest this when going over 4000 12-11-11 . This behavious will be manifested on other brands of motherboard so it is not specific to Maximus VIII series.

You will first notice a long wait at 79 code then another long wait at b4 ,culminating with another long wait at 99 then finally pushing into a weird error code regarding Mei before finally entering OS.

 

 

 

 

Also this is related to not being able to shutdown or restart from OS being stuck at 04 ore 05 code .

As weird as this may sound, a stick that had this behavior at 3733 12-12-12 could pass DDR4-4080 12-11-11 with copy-wazza so on current platforms there is no connection between this behavior and benching potential.

Cooling the memory to -50 or lower got me rid of this behavior but overclocking potential did not get better sadly.

 

2. Voltage tolerance behavior

4F is the debug code on B-die for expressing that you passed the voltage tolerance limit on your sticks.

 

For example on most of the sticks I could bench relative stable 0.03-0.04 under this limit . When you reach the “Danger Zone†the stick will let you know it is not confortable by either getting weird blue-screens or by receiving “NOT CONVERGENT IN SQRT†error in SuperPI 32M.

 

3. Different behaviours under stress

- NOT CONVERGENT IN SQRT while running 32M,either problem with voltage tolerance or TRCD too tight

-exit benchmark while running 32M, either too tight RTL or too low VCCIO, 1.35V and -100 will definatly maxxx-out your IMC potential.

-errors in DJIKSTRA test of Geekbench , try lower-higher voltage than 32M .

-XTU errors . In my case XTU required higher TRFC for complete stability than 32M,however XTU also benefits lower trfc to 200-220 so the only way of having complete stability for bugging the ultimate score was cooling mems to 0-10 degrees.

 

 

***

Edited September 23, 2016 by Alex@ro

[sabishiihito]

Any tips for running B-die 4000+ on air?

 

Sent from my SM-N930T using Tapatalk

[Alex@ro]

Well these is on air mostly . U need Impact + Decent IMC + proper sticks.But for 4000 at least good sticks is what you need,4133 is more tricky with board and IMC.

[sabishiihito]

That super slow boot is annoying as all getup, is cold the only fix?

 

Sent from my SM-N930T using Tapatalk

[Alex@ro]

As far as i found yes.

[Alan_Alberino]

Any tips for running B-die 4000+ on air?

 

Sent from my SM-N930T using Tapatalk

 

In Asrock, I set 3866 c12-12-12-28 240 with rest on auto, then tight secondary and tertiary timings at 3866 and then push to +4000 and make the final profile. I tested 3 different CPUs and all made 4133 12-11, good board is more important than good IMC

[MaRsXzU]

What are the relations between tWTR_L i tWTR_S ? And similarly tRRD_L i tRRD_S? Shoud I keep any distance between of them? Im working on Hynix sticks.