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

Associate
Joined
29 Jun 2009
Posts
549
This is just a question borne out of curiosity more than anything else but, everything else being equal, are 2x16GB kits easier or harder to overclock than 2x8GB kits (or should they be the same)? Assume speed, latency/timings etc are the same.
Yes, definitely harder. You tend to have to slacken TRDCRD by 1 and sometimes TRRD's need to be at 4,6 instead of 4,4. I've seen double sided sticks need twtrs settings at 5,14 as well. Definitely harder...but you do gain around 5% memory performance so sometimes its worth it.

MrPils tRCD14 and tRP11 settings were only to see if they would run and how they go in this bench and I would say far from stable.
Voltages used were
Vdimm 1.52
Soc 1.0625
Vddg ccd 1.0
Vddg iod 1.0
Vddp .95
ProcODT 28.2
MemCadBus 30 20 20 24
For stability testing my setup so far seems to need tRCD15 and CR2 with GDM off. Have passed TM5 with tRP12 but will test more tonight.
Also when looking to run tRCD14 , 15, 16 etc is this more the cpu than the ram sticks? Was looking to try another itx mb too for the 2 dimm slots

Thanks so much for that, really useful info :)

I was running procodt 28.2 on my MSI as it was the best setting, but this Gigabyte board doesn't go that low it only goes to minimum 30 which is no good. My second best setting was 53.3ohms so I'm having to run that now. Not sure how much overall effect that has, but 53.3 has a significant effect reducing errors in memtest when on the edge of stability. Wish I could try 28.8 as that was clearly better on my MSI.

Your timings and results inspired me to do some investigation last night. My results may help...

tWRRD 2 - Massive boost in stability for tRDCRD but you take a performance hit.
tRDWR 8 - I cannot post at this setting over 3600mhz. Black screen and oc failed. Tried in combination with tWRRD 1 @ 3733mhz 1:1 and it played an awesome tune through the internal speaker and ate my bios (requiring a flashback).
TRCDRD 14 - I can run memtest default bench without error on 2 out of every 3 runs when using tWRRD 2 and ProcODT 53.3 @ 1.52vddr. SOC voltage didnt help, didnt try adding IOD voltage though it got too late. Will have a quick try at that tonight, but initially the performance was identical in memtest compared to TRCDRD 15 and tWRRD 1. TRCDRD 14 and tWRRD 1 generates a ton of errors, sometimes even on the easy test.

Looking at the timings in your recent membench screenshots you should be able to drop tCWL to 12 with no stability hit or additional voltage requirement. tRDWR 9 and tWRRD 1 may give you better stability and performance balance than tRDWR 8 and tWRRD 2. I think your 2T stability requirement is probably coming from those two timings (based purely on the fact that GDM and 2T mode gain me absolutely nothing at all in terms of timings or speed and thsoe are the only two things we are doing differently).

The only other thing worth looking at is maybe trying to drop your VDDG CCD. It drops your core temps a small amount under intense loads like prime. I found that you can just tweak it down slowly until you get a no post, then go back up a notch. Mine no posts at 0.775v and is solid stable at 0.8v. Compared to running the voltages in sync its worth 2c to 3c in prime 95 blend. IOD and VDDP are processor specific but usually like to be around 100mv apart.

I did have a really odd and repeatable effect when testing TRCDRD 14 last night using procodt 32ohms - my bandwidth result running membench went up significantly. Errors went through the roof too, so it was probably due to instability rather than an actual increase but I need to test this with my stable settings tonight to verify.

I also need to test GDM and 2T mode to see if that will allow tRDWR 8 or primary reductions with proper stability and if so performance compare against 1T. Maybe not tonight though, ended up playing til 3am last night by mistake...feeling it today lol.
 
Last edited:
Associate
Joined
29 Jun 2009
Posts
549
Managed to test GDM and CMD 2T quickly while having lunch, did not allow either TRCDRD reduction or tRDWR 8, no change in behaviour at all.

Also tested 32ohms ProcODT. No out of line behaviour at all with my stable timings and completes the test in normal time. I retested again as a sanity check and it still shows a weird 10gb/s increase in write speed when trying to run membench @ 32ohms with TRCD 14 along with being WAY more unstable than any other ProcODT settings I tried. I'm guessing corruption in the memory is artificially inflating the write score and I'm going to ignore it from now on.

Also realised that with being able to run tRDRDSC_L and tWRWRSC_L at 2 you should also be able to run tWRWRSD and tWRWRDD at 4...and you can. Quick 100% pass completed fine but needs a more thorough test. With that said the recommendation for setting tWRWRSD and tWRWRDD can probably be changed to "Optimal and Minimum = 2 x tRDRDSC_L / 2 x tWRWRSC_L, increase to reduce performance and gain stability".

Performance increase from setting the above timings to 2/2/4/4 from 3/3/6/6 is -0.4ns Custom latency, -0.1ns Random Latency and around 0.33s reduction in membench default test. Stability and voltage requirements seem so far unaffected.

Updated timing recommendations behind spoiler tag below.

B-Die Single Rank Timings for 1:1

Primary Timings:
Recommend beginning at 16-16-16-32 1.5v with Fan and 1.4v without fan. Tune timings down once stability has been tested.
TRCDRW = CL / 2
TRAS = TRCDRD+CL but can go lower on B-Die at the risk of data corruption if unstable.

Variable Secondary Timings:
TRC=TRAS+TRP
TRFC=6xTRC (TRFC has a hard limit which increases with frequency. Causes data corruption if unstable)
tCWL=CL-2 @ CMD 1T / CL-3 @ CMD GDM / CL-4 @ CMD 2T

Fixed Secondary Timings:
tWR 10
tRTP 5 (tWR/2)
tWTR_S 4
tWTR_L 8
tRRD_S 4
tRRD_L 4
tFAW 16
tCKE 0 or 1 - should have no effect with dram power down disabled

Tertiary Timings:
tRDRDSC_L 2 (May need 3 on older generation boards)
tRDRDSC 1
tRDRDSD 4 (Optimal and Minimum = 2 x tRDRDSC_L, increase to reduce performance and gain stability)
tRDRDDD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tWRWRSC_L 2 (May need 3 on older generation boards)
tWRWRSC 1
tWRWRSD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tWRWRDD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tRDWR Auto - Target 8/9/10 (Strictly bound to memory quality. Tune down from Auto once stable if possible. Try 8/2, 9/1, 6/4, 7/4)
tWRRD Auto - Target 1/2/4/7 (used in tandem with tRDWR, adds a latency buffer when switching from read to write to write to read)
 
Last edited:
Associate
Joined
11 Dec 2016
Posts
2,010
Location
Oxford
Sort of in topic
Ryzen memory scaling in games review by Guru3d
https://www.guru3d.com/articles-pages/amd-ryzen-ram-scaling-effect-in-games,9.html
index.php


Both dual rank (and 4x8) configurations perform admirably at 3200 CL14
4x16GB 3200 CL14 beating 2x8 3733 CL14 IF 1:1 in all games.
I had no idea Ryzen likes dual rank so much.
 
Soldato
Joined
15 Oct 2019
Posts
11,656
Location
Uk
Finally managed to break 200 with my 8pack 3200/14 kit.

1988.png


Interestingly unlike the 3900XTs gains there is quite a large performance regression with SMT disabled on a 3600.

htdisabled.png
 
Last edited:
Associate
Joined
28 Sep 2018
Posts
2,242
Sort of in topic
Ryzen memory scaling in games review by Guru3d
https://www.guru3d.com/articles-pages/amd-ryzen-ram-scaling-effect-in-games,9.html
index.php


Both dual rank (and 4x8) configurations perform admirably at 3200 CL14
4x16GB 3200 CL14 beating 2x8 3733 CL14 IF 1:1 in all games.
I had no idea Ryzen likes dual rank so much.

It’s not that Ryzen specifically likes dual rank. It’s that rank interleaving is very strong for performance. That’s a memory architecture advantage and not cpu related. Same on Intel.

Got a buddy on comet lake doing 16/4400 on dual rank bdie. Daily stable. That’s fast!
 
Associate
Joined
11 Dec 2016
Posts
2,010
Location
Oxford
It’s not that Ryzen specifically likes dual rank. It’s that rank interleaving is very strong for performance. That’s a memory architecture advantage and not cpu related. Same on Intel.
There are not many such tests with dual rank and always done with Intel.
Apprently Ryzen memory controller can take advantage of 8 ranks, this is the first time where I see 4x16GB have consistent advantage over 4x8 and 2x16GB.
 
Associate
Joined
29 Jun 2009
Posts
549
Finally managed to break 200 with my 8pack 3200/14 kit.


Interestingly unlike the 3900XTs gains there is quite a large performance regression with SMT disabled on a 3600.

Well done dude, nice result. Congrats on reaching your goal too, its always good to actually achieve something you set out to and find a little more on top :)

Can you run GDM disabled with trcdrd 15 and twrrd 9? If not you haven't got much left to tune down there. TRC can go one lower (and therefore trfc 6 lower) with your current primaries and TRTP can come down to 5 but neither of those will rock your world and I don't see anything else that wouldn't need a big boot of voltage.

I'm beginning to wonder if these 8 pack sticks are optimised in some way for GDM and 2T, sacrificing some 1T ability along the way. Or maybe the Patriots are optimised for 1T with all those extra capacitors they have on the PCB and they are sacrificing GDM/2T ability? Honestly if the stability improvement allows TRCDRD 14 and TWRRD 8 with GDM enabled then that's probably worth it, I would expect that to be faster than a 1T setup at TRCDRD 15 and TWRRD 9. GDM and 2T gain me nothing in real stability with these Patriots yet 1T is decidedly easy.

There are not many such tests with dual rank and always done with Intel.
Apprently Ryzen memory controller can take advantage of 8 ranks, this is the first time where I see 4x16GB have consistent advantage over 4x8 and 2x16GB.
Its about as big a boost as you get from going from dual to quad channel (+2.5%) on x99. That's pretty weird...I wonder what its doing? They're only using a 3700x so its not like it has two CCD's to do anything funky with.
 
Soldato
Joined
15 Oct 2019
Posts
11,656
Location
Uk
Well done dude, nice result. Congrats on reaching your goal too, its always good to actually achieve something you set out to and find a little more on top :)

Can you run GDM disabled with trcdrd 15 and twrrd 9? If not you haven't got much left to tune down there. TRC can go one lower (and therefore trfc 6 lower) with your current primaries and TRTP can come down to 5 but neither of those will rock your world and I don't see anything else that wouldn't need a big boot of voltage.

I'm beginning to wonder if these 8 pack sticks are optimised in some way for GDM and 2T, sacrificing some 1T ability along the way. Or maybe the Patriots are optimised for 1T with all those extra capacitors they have on the PCB and they are sacrificing GDM/2T ability? Honestly if the stability improvement allows TRCDRD 14 and TWRRD 8 with GDM enabled then that's probably worth it, I would expect that to be faster than a 1T setup at TRCDRD 15 and TWRRD 9. GDM and 2T gain me nothing in real stability with these Patriots yet 1T is decidedly easy.
It don't let me disable GDM unless I reduce the speed to 3600 or run 16s at 3800. I think your right about being at the limits of what I can get out of these as I'm already at 1.49v and don't really want to go higher but still pretty pleased given these are not the higher end 3600/14 kit.

Not sure if it's because of GDM but the SCLs stick at 2 on my B450.
 
Associate
Joined
29 Jun 2009
Posts
549
It don't let me disable GDM unless I reduce the speed to 3600 or run 16s at 3800. I think your right about being at the limits of what I can get out of these as I'm already at 1.49v and don't really want to go higher but still pretty pleased given these are not the higher end 3600/14 kit.

Not sure if it's because of GDM but the SCLs stick at 2 on my B450.

Yeah you've done well there from a 3200 c14 kit for sure. I'd stay below that 1.5v limit if you can, you could probably convince yourself up to about 1.52v for a particularly tempting timing decrease but honestly its not likely noticeable ever. Once you get membench below 200s your system is flying and the bottlenecks will usually be elsewhere.

Thanks for the info on your B450 and the SCLs. I wonder if its something particular to the MSI B450i then (that's the only B450 board I've used) or maybe case by case between vendors and boards? It would make sense if the ability to run the SCLs at 2 was something which was developed during the 400 series and therefore just works on the 500 series. Still, its a nice little latency boost that's for sure. AIDA likes it :)
 
Associate
Joined
9 Oct 2013
Posts
51
Thanks for all the info you have put out in the thread MrPils. It has been very helpful.

Today I spent more time trying to pass TM5 at tRCD14 but not to be.

Settled on the below.
Bios settings
Vdimm 1.475
Soc. 1.08125 droops to just over 1.05
Vddg ccd .975
Vddg iod. 1.0
Vddp. .925

y1viqJC.jpg.png
 
Last edited:
Associate
Joined
29 Jun 2009
Posts
549
Thanks for all the info you have put out in the thread MrPils.
Today I spent more time trying to pass TM5 at tRCD14 but not to be.
Settled on the below.
Bios setting
Vdimm 1.475
Soc. 1.08125 droops to just over 1.05
Vddg ccd .975
Vddg iod. 1.0
Vddp. .925

You're very welcome dude! You've got an awesome setup there, time to sit back and take it for granted now... Honestly though memory tuning is dull and repetitive, I find it really helps to have a bit of company along the way even if its only an active forum post :). I mean I've barely touched my memory setup in nearly a year and thanks to the back and forth in this thread I've found an improvement I never would've tried otherwise. I still wish I could get tWRRD to run at 8, it's the carrot that's just out of reach at the moment but I think it's going to have to remain that way...

Just a final bit of advice...if you get any weird behaviour in games stick the bankgroupswap options to disabled. I've had some very odd stuttery behaviour with BGS options enabled when playing (some) actual games instead of benchmarks...

If I get a few hours at the weekend I'll do some 2x16gb testing as I never spent the time to figure out the subtimings for dual ranks. If the stars do align and it actually happens I'll post the info up here for any of you using the 2 x 16gb kits :)
 
Associate
Joined
29 Jun 2009
Posts
549
Does anyone have some tried and tested settings to get the 2 x 16GB 3600MHz CL16 kit to 3733MHz CL14?
Not optimised no, but ill be starting at this and expecting it to be stable:

Primary Timings:
16-16-16-32-1T 1.5v with Fan / 16-17-16-33-1T 1.4v without fan. Tune timings down once stability has been tested.
TRCDRW = CL / 2
TRAS = TRCDRD+CL but can go lower on B-Die at the risk of data corruption if unstable.

Variable Secondary Timings:
TRC=TRAS+TRP
TRFC=6xTRC (TRFC has a hard limit which increases with frequency. Causes data corruption if unstable)
tCWL=CL-2 @ CMD 1T / CL-3 @ CMD GDM / CL-4 @ CMD 2T

Fixed Secondary Timings:
tWR 20
tRTP 10 (tWR/2)
tWTR_S 5
tWTR_L 14
tRRD_S 4
tRRD_L 6
tFAW 24
tCKE 0 or 1 - should have no effect with dram power down disabled

Tertiary Timings:
tRDRDSC_L 2 (May need 3 on older generation boards)
tRDRDSC 1
tRDRDSD 4 (Optimal and Minimum = 2 x tRDRDSC_L, increase to reduce performance and gain stability)
tRDRDDD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tWRWRSC_L 2 (May need 3 on older generation boards)
tWRWRSC 1
tWRWRSD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tWRWRDD 4 (Optimal and Minimum = 2 x tWRWRSC_L, increase to reduce performance and gain stability)
tRDWR Auto - Target 8/9/10 (Strictly bound to memory quality. Tune down from Auto once stable if possible. Try 8/2, 9/1, 6/4, 7/4)
tWRRD Auto - Target 1/2/4/7 (used in tandem with tRDWR, adds a latency buffer when switching from read to write to write to read)

Tertiary timings start at auto and tweak down to finalise. You'll get an idea pretty quickly whether theyll go as low as single rank settings. Id expect to run SCL's at 3 and SD/DD at 6 but really good sets may do 2 and 4. tWR and tRTP should come way down, but I've seen bios issues where a high tWR was required with dual rank sticks before so start high and tune down quickly. 14-16-13-30-1T is my target at around 1.45v...but secretly hoping for 14-15-13-29-1T @ around 1.5v (not confident lol).

If you do have a go at it and you wouldn't mind sharing your notes that may save me some time at the weekend :)
 
Soldato
Joined
27 Jul 2007
Posts
6,085
Wow, that's comprehensive. Thanks!

I'll be honest, I'm not looking to eek out every last morsel of performance because I just don't have the time. 3733MHz CL14 with stable settings would do me just fine.
 
Associate
Joined
29 Jun 2009
Posts
549
Wow, that's comprehensive. Thanks!

I'll be honest, I'm not looking to eek out every last morsel of performance because I just don't have the time. 3733MHz CL14 with stable settings would do me just fine.

Try the above but leave the Tertiary timings on auto and don't bother tuning down any Secondary timings if you don't want the last few percent. TRCDRD will either be 16 or 17 depending on your voltage limit and kit quality, CL14 will either work or not depending on voltage limit and kit quality. Don't go much over 1.4v without direct airflow on the memory, you'll get random errors if/when your overall system gets hot. Just use the 16-16-16-32 or 16-17-16-33 main timings to make sure you can run 3733mhz stable, then tweak down your CL from there.

You should really tune your SoC, vDDG IOD/CCD and your vDDP voltages to run at 3733mhz as well if you haven't already. Auto voltages usually work, but massively overvolt in nearly all cases and take power budget from your CPU cores (therefore reducing your core clocks). You should tune voltages down from auto after you have your memory timings and voltage dialled in...faster memory stresses the Infinity Fabric more so voltage requirements may go up as you increase your memory performance making it pointless to do this before memory tuning. If you had done this before but have now upgraded from 2x8gb to 2x16gb you'll need to do it again anyway as simply running dual rank is harder on the Fabric so it can sometimes need a bump more voltage compared to single rank.

Voltage cheat sheet:
1) vDDG IOD is the main voltage you will need to raise for Fabric stability. Start at 1150mv (AMD Auto for 1866mhz Fabric @ 1:1) and work down, I would pay serious consideration as to whether any speed gains are worth running this voltage above 1050mv on a daily basis. Default is 700mv @ 2133mhz, 900mv @ 3200mhz and 1200mv :eek: @ 3800mhz.
1) vSOC should always be below 1.25v and must be at least 50mv (preferably 100mv) above vDDG IOD. Raise SoC as you need to raise vDDG IOD, IOD voltage requirement will likely govern your final SoC voltage.
2) vDDP should be stable around 100mv below vDDG IOD (but not over 950mv)
3) vDDG CCD is stable at 100mv less than vDDP on all 4 of my Ryzen 3000 cpus.

Depending on your standards of stability, run memtest for either 400% single coverage, 800% single coverage or 1200% single coverage for testing your final settings. Takes a while on 32gb...


If you don't want to do it that way you can bang in random numbers and voltages if you want, but I guarantee it'll take a lot longer... I suppose the easiest possible way to your target would be to set XMP enabled, leave everything on auto and just try increasing to 3733mhz with 1866mhz fclk and memory/uclk at 1:1. If that doesn't work try TRDCRD at 17. Once stable drop CL and test stability again. Performance will be crap compared to what your kit/system would be capable of, but its quick and easy. It may turn out to be slower than 3600mhz xmp depending on your auto voltages and the resulting cpu core clock loss, but the memory subsystem will definitely be faster than 3600mhz XMP.
 
Back
Top Bottom