Zen 2 Quest (3733 CL14-15-15-28 @ 1.47v Get)

[Bloax]

Feb 2020 Edit: Despite wild hardware adventures and a board swap for the better, pretty much every drop has been squished by now
Furthermore, I'd like to remind people to test your **** ****; 3733 had a performance regression on the b450 Tomahawk compared to 3666, so the premise of this thread is rather funny in retrospect.

Aug 3 2019 Edit: 1T command rate @ 1.47v and 0.885 VDDP almost almost stable
Aug 5th Edit: There's no discernable difference in performance between 1T and 2T in Overwatch, so I'll take the free stability.

As some here may have noticed, I've had a good ol' time with Zen 2
As some may also have noticed, people have gotten awfully softOCPost-y - people claim CPU overclocking is dead this, 3600 CL16 that

Here's to a toast for the good old times, when [H]ard was [H]ard and squishing performance was just a good old pastime.

The board was acting proper trash at first. Did that stop me? Nah, I tried again once it got updated.
Turns out 3600 CL14 worked just fine, but I didn't entirely trust the board.
Did that stop me though?

No - I went ahead, scribbled a couple of pages of notes in a paper notebook, and got to work.
Some CMOS resets and full config re-entering sessions (god bless paper, huh) later, we got to 3733 CL14ish running full retard 3533 subtimings (1900 fClk POSTed once on this 3600, but it had a hissyfit) - courtesy of this magnificent piece of work and lookup tool https://www.techpowerup.com/review/amd-ryzen-memory-tweaking-overclocking-guide/

Most of the stability thanks to lowering VDDP from 0.900(4) to 0.890 and raising VPP to 2.57, the former of which is actually visible in Ryzen master (thank ****)

But we weren't entirely stable - every hour or so a proper memory test would throw a single error (or two in test #6 for memtest86)
Is 99.9% stable good enough for the main purpose of cranked to **** memory; high refresh gaming?

29-07-2019, 04:00:48 Overwatch.exe benchmark completed, 48779 frames rendered in 244.656 s (shadows off, lighting medium, PBO+AutoOC 3600 CL14, unimpressive subtimings)
Average framerate : 199.3 FPS
Minimum framerate : 186.1 FPS
Maximum framerate : 200.0 FPS
1% low framerate : 175.1 FPS
0.1% low framerate : 131.1 FPS

30-07-2019, 13:44:37 Overwatch.exe benchmark completed, 48937 frames rendered in 245.203 s (shadows off, lighting medium, 3733 CL14-15-15-30 good subtimings)
Average framerate : 199.5 FPS
Minimum framerate : 188.6 FPS
Maximum framerate : 200.0 FPS
1% low framerate : 183.8 FPS
0.1% low framerate : 144.1 FPS

Yeah, it actually is - I only ever got a single crash in about twenty hours of PC usage.
But then I discovered another BIOS update - now featuring exporting/importing setings via USB, so I decided to try and try again.
Some tinkering later (and a failed attempt to get 1T more than just 95% "hey if it only errors every 6 minutes it's fine" Bench Stable) with minor sub-subtiming improvements..

And here we are, a final procODT 48->53.1 - something which I didn't think to try due to attempting it on the initial BIOS, but turns out this board can handle that - and a drop from 1.51v to 1.48v

3733 CL14 (hey if G.Skill can get away with calling their Exotic kit 3800 CL14 @ a chubby CL14-16-16-36, 14-15-15-28 definitely qualifies) 99.999% stable at 1.48v on a 3600 that wouldn't do 1900 fClk

Cheers, and may the Last Bastion of CPU Overclocking - [H]ard memory overclocking - smile upon you.

 

[Strelok]

I was getting really frustrated with memory OC'ing because my scores kept getting worse/staying the same even though I went from CL16 (16-16-18-36) to CL14 (14-14-17-30). Funnily enough, I switched Geardown and Powerdown to auto and everything was fixed. I'm assuming it wasa the powerdown setting, since geardown is enabled. That said, I didn't notice a huge difference in the games I tested, probably because I was heavily GPU limited anyways.

Here are my settings (32GB e-die)

 

[Bloax]

Well yes - memory is most of all a limiting factor in High Refresh minimum framerate scenarios.
Which is rather relevant for me seeing that I have my sights on a Good Motion Clarity Monitor (tm) sooner rather than later to conclude a two-year long SteadyMouse experiment.

To my surprise, 1T works but seems just barely more stable than my previous 2T setup prior to upping procODT and lowering voltage;

So more than likely full 5-hour memory test stability would be achieved with +/- 0.005v, but unfortunately there's no such granularity on this board.

Don't forget; never give up.

 

[Dan_D]

I'm going to try and do some more performance tuning with Ryzen and RAM myself. As i said in our review of the 3900X, overclocking in the traditional sense may be on its way out, but there are still things to tune for more performance.

I'm going through various modules I have to see what they can do.

 

[Bloax]

The good news is that we're almost there in regards to getting full stability @ 1T

the bad news is that the sweetspot is likely 1.465v or 1.475v, which isn't Get-able with 0.01v granularity

And crazy-town memory tweaking isn't even much of a question in Zen 2 land; it may not show up in synthetics too much, nor may it show up in average FPS too much;

But it does show up in regards to turning your 0.1% lows into stock 1% lows
(quick comparison between 3733 CL14 squished to **** vs. stock 3200 CL14 XMP)
overwatch is a pretty good RAM benchmark these days, now that we have replays which are Practically Comparable (we tested!) to actual gameplay performance while being 100% consistent between runs


Based on my experience with my first Ryzen system being this Zen 2 system - CPU overclocking has gone from simple to crazytown.
Why bother yourself with silly concepts like vCore, when you can be worrying about VRAM voltage, vSOC, VDDP, VPP, the highest procODT the board can squish out of your memory config (just because it works on 2 DIMMs doesn't mean it will on 4!!!), raising fClk to your CPU's maximum if your memory refuses to budge past 1700 uClk ...
and that's without touching timings!

I'd say it's alive and well yes sir

 

[JosiahBradley]

How'd you get overwatch to benchmark? Did I just miss an update? I'd love to test my RAM out.

Running 4×8GB at 3800 CL 14-16-16-3* at 1.5v with a 3950x FCLK 1900. I want to push it further and get tighter timings. Even the ryzen dram calculator can't help me here as it doesn't understand this high setting apparently.

 

[stickx911]

Over clocking is alive on these only because the infinity fabric relies on ram speed. Thus, oc’d ram equals oc’d cpu? Or am I missing something?

All I’ve seen is hours spent trying to get 100 extra mhz on a 4ghz cpu, and that gain normally comes at the cost of single thread or a significant hit on efficiency. The auto OC on the cpu hits near max on it’s own so there’s so little to gain doing multiplier or bclock adjustments. It’s not like getting 5ghz out of a 2500k or pencil modding an Athlon.

Not to say that tweaking isn’t available, and for me, I still play with the adjustments because that’s more fun than playing the games. The rewards just aren’t nearly as sweet. Took cpu makers 20 years to finally build in a stable and quality auto OC system.

 

[Dan_D]

Over clocking is alive on these only because the infinity fabric relies on ram speed. Thus, oc’d ram equals oc’d cpu? Or am I missing something?

All I’ve seen is hours spent trying to get 100 extra mhz on a 4ghz cpu, and that gain normally comes at the cost of single thread or a significant hit on efficiency. The auto OC on the cpu hits near max on it’s own so there’s so little to gain doing multiplier or bclock adjustments. It’s not like getting 5ghz out of a 2500k or pencil modding an Athlon.

Not to say that tweaking isn’t available, and for me, I still play with the adjustments because that’s more fun than playing the games. The rewards just aren’t nearly as sweet. Took cpu makers 20 years to finally build in a stable and quality auto OC system.

To a point. The Zen architecture in general likes memory speed and lower latencies. The Ryzen 3000 series has RAM speed tied to the FCLK and as a result, the two values are linked. So up to 3733MHz, the Ryzen 3000 benefits from both clock speed and lower latencies. However, at 3800MHz+, the FCLK is set to 1800MHz and performance actually gets worse. However, in theory you can compensate for that if you can get the clocks high enough to overcome the lower FLCK.

AutoOC or PB2 gets you the best overall results on the platform when combined with low latency RAM and decent RAM clocks. Tuning is very much alive on these things albeit overclocking in the traditional sense isn't. Which is why I had a lot more fun with the 10980XE than I did with the 3950X in that department.

 

[Bloax]

now that i got back into this account after forgetting the password and email for a month or two

How'd you get overwatch to benchmark? Did I just miss an update? I'd love to test my RAM out.

Running 4×8GB at 3800 CL 14-16-16-3* at 1.5v with a 3950x FCLK 1900. I want to push it further and get tighter timings. Even the ryzen dram calculator can't help me here as it doesn't understand this high setting apparently.

Overwatch has had replays for about half a year or a bit more now, which behave very closely to The Real Deal, and which are mostly deterministic in their nature. Thus you can consistently benchmark e.g. a tense match of King's Row or what have you.
For capturing the performance data, you can use CapFrameX

Though I will give a little warning about software performance capturing; the frametimes were identical between me running a 3700x (the now two-pin-less) with SMT on and SMT off, yet the game had very noticeable microstutter with SMT off that I could physically see on this 240 Hz monitor.
Yet if you were to trust frametime data, there was no difference at all!
So for real conclusions, you always have to give it proper respect and actually sit down and try your stuff out live.


Generally, then I would advice people to stay off attempting 1900 FCLK - ever since I upgraded to this x570 Phantom ITX, which has vastly better power filtering than the b450 Tomahawk I was on previously - I've been able to run 1900 FCLK.
But see, running doesn't mean that it's 100% stable and useful; 1866 FCLK ran on the b450 Tomahawk, but provided a significant (~10%) performance degradation over 1833. It's important to test and confirm performance improvements over 1800 FCLK, which are usually rather linear.
However, the moment the gains become non-linear, or barely existent as it would be, is when you can tell that stability is down the ****ter.
In the case of this new board, performance increases between 1800, 1833 and 1866 - but barely changes going up to 1900 from 1866.
And sure enough, the system crashes doing intensive workloads when it's not the only thing drawing power in the house.

My personal memory settings have improved a bunch since this initial post, and the sticks are now chilling at 1.57v with a fan zip-tied next to them (picture is prior to zip-tie)

I have a sneaking suspicion that rdrd/wrwr SCL + Sd/Dd's go down so low (SCL 2 + Sd/Dd 3 had ~3 hour memory error intervals despite a tRFC this low) because this is a 1-CCD CPU with a halved memory write interface, based on how I've had the write speeds be a full 14.5% faster than Half-of-Read with some semblance of stability.


As for FCLK stability testing, I can recommend

youtube-dl.exe https://www.youtube.com/watch?v=OTRJxPPbPCM

Downloading a very long video with youtube-dl, then transcoding it with x264 Medium using a few simple batch files

@echo off
cd /d %~dp0
"C:\StabilityTest\x264.exe" --preset Medium --crf 17 --frames 36000 -o quickTest.mkv "Wizards & Warriors (PC) - Session 4-OTRJxPPbPCM.mp4"
echo Howdy partner, we seem to be done?
pause

@echo off
cd /d %~dp0
x264.exe --preset Medium --crf 17 -o Test.mkv "Wizards & Warriors (PC) - Session 4-OTRJxPPbPCM.mp4"
echo If we have a full 6 hours, 34 minutes and 23 seconds of video in our test MKV then that's a full pass!!
pause

Over clocking is alive on these only because the infinity fabric relies on ram speed. Thus, oc’d ram equals oc’d cpu? Or am I missing something?

Zen is a very data-hungry architecture, and it only seems to get worse and worse as time progresses (Zen 2 is much faster) - I've seen very significant gains just from cranking the **** outta memory in tasks such as x264 encoding;

1833 (fast timings) fclk: 166.02 166.66 166.56 fps
1833 (TURBO)        181.62 fps 181.22 fps

Even when just squashing secondary/tertiary timings compared to already fairly-ok Ryzen DRAM Calculator settings, you can see performance improvements in vidyagames;

1867 squished timings

avg: 240.0 240.0 240.0 239.7 239.3
p5: 205.7 204.9 205.2 203.7 203.0 (204.5 avg) +0.49%
1%: 168.0 168.6 168.0 164.6 165.6 (167) +1.34%
p02: 162.9 164.4 164.9 158.7 161.3 (162.4) +1.44%
0.1% 117.6 121.4 120.9 117.8 119.7 (119.5) +2.75%


1867 dram calc settings (same trfc):

avg: 239.5 239.2 239.7 239.7
p5: 203.0 202.4 203.9 204.6 (203.5)
1%: 160.1 164.8 167.0 167.3 (164.8)
p02: 152.8 160.6 162.6 164.3 (160.1)
0.1% 110.7 118.8 117.2 118.6 (116.3)

And as stated in my notes, the tRFC - a setting very impactful to performance, as it determines how long the memory spends refreshing its contents - was kept the same at 232, a 22% drop from the setting recommended in DRAM Calculator.
(this drop being possible because I'm running my memory at 1.57v instead of ~1.45v)

So yes, there's lots and lots of performance to be had from squishing every last drop out of your b-die (easily identifiable by Frequency / tRCDRD >= 200 memory bins)
a process which I've slowly worked away at every now and then in the evening, continuously napping and running memtesthelper in safemode.

Cheers, and have fun

 

[Eduardo_Domingot]

You have really been a great source of info on this Bloax with all the videos and personal experience you have provided. I am still trying to lower timings and I think I might need to drop to 3666 in order to get their.

 

[Bloax]

The difference between 3666 and 3733 is not exactly as notable as 3600 vs. 3800 - so that's no biggie.
Apparently I'm hearing rumblings of Renoir going up to 2000 fclk,
so that may be an indicator of where Zen 3 (well, the trajectory from Zen1->+->2 gave it away) is likely to go.

That's gonna be uhh an adventure to crank back down again.

now to decide whether to dual-rank Sammy B or not

 

[Eduardo_Domingot]

The difference between 3666 and 3733 is not exactly as notable as 3600 vs. 3800 - so that's no biggie.
Apparently I'm hearing rumblings of Renoir going up to 2000 fclk,
so that may be an indicator of where Zen 3 (well, the trajectory from Zen1->+->2 gave it away) is likely to go.

That's gonna be uhh an adventure to crank back down again.

now to decide whether to dual-rank Sammy B or not

Yeah that 2000 fclk will be nice, I doubt it will surpass Intel in gaming perf but it should bring down latency a decent amount.

 

[Bloax]

I found a pretty nice deal on some snazzy Gizzy Trizzy Z 2x16 3200-14-14-14

So I expect my weekend plans for next weekend to be a bunch of tomfoolery with that stuff.

Personally, then I'm not particularly interested in the ""GAMING PERFORMANCE"" crown dick-measuring contest;

Zen2 is completely capable of delivering rock solid, no-drops 120Hz gameplay given a little memory massaging
Refreshlake is completely capable of delivering rock solid, PoTenTiALLy no-drops 120Hz gameplay given decent memory and not having any security patches forced down your throat.

It doesn't matter which one of them is 5% better here, 8% better there - NONE of them deliver no-drops (and especially, no-fuss) 240 Hz gameplay.
The best display you can reasonably get (realistically, on a long-term payoff plan) right now is an LG C9 or CX OLED TV, which is a 120 Hz panel.
Either will drive that thing, but Zen 2 is the safer bet as far as "lack of performance degradation over time" goes.

 

[Bloax]

This stuff arrived faster than anticipated, and what I didn't expect is that the "weird zen2 b-die trick" would work on dual-rank;

Yet here we are, with tRAS 21 tRC 29 once again.
Timings are by no means final yet, nor have I even confirmed their long-term stability (as I am yet to sleep)


Edit: Oh right, I forgot to mention that these things weigh a ton - and the heatsinks are fairly hefty and don't even seem to be fashion accessories. Neat.

 

[Bloax]

Well well, now there's a status update for you;

After a bunch of initial tinkering;
such as establishing tRFC scaling from 1.46v to 1.56v to not scaling further past that, which was 252 error-fest at 1.46v, 228 error-fest at 1.56v; ergo about 260 stable @ 1.46v (suspiciously good), 232/234 stable @ 1.56v, 228 not budging at 1.6v -
uncovering that tRAS 21/tRC 29 still work, and bumbling right into tertiary walls (hmm???) with specifically tWRWR Sd/Dd blowing up (not POSTing) at 5 but otherwise everything was peachy thus far

I decided to just go for a "money shot" - sarcastically referring to a "low blow" meant to end the tinkering session, and finally go to The Big Sleep for validation of Current Progress.
Whacked in

GDM Off(, 1T), tCWL 16->13, tCL 16->14, tCL 16->14, tRCDRD 16->14, tRP 16->14

wow what a surprise it doesn't POST loool

alright, one last one just for ****s and giggles;

tWTRL 12->10, tWR 12->10, tCL 16->14, tRCDRD 16->14, tRP 16->14

To my surprise, it POSTed!

it was even making progress in the dram calc bench without spazzing out!

You know, this was getting a little funny.

you know, I didn't plan for it to instantly succeed.
I thought these dual-rank thingies were supposed to have primary/trfc deficits compared to single-ranks, not out-perform them!

Now I just have various little subtimings to push, bah!

 

[Eduardo_Domingot]

Well well, now there's a status update for you;

After a bunch of initial tinkering;
such as establishing tRFC scaling from 1.46v to 1.56v to not scaling further past that, which was 252 error-fest at 1.46v, 228 error-fest at 1.56v; ergo about 260 stable @ 1.46v (suspiciously good), 232/234 stable @ 1.56v, 228 not budging at 1.6v -
uncovering that tRAS 21/tRC 29 still work, and bumbling right into tertiary walls (hmm???) with specifically tWRWR Sd/Dd blowing up (not POSTing) at 5 but otherwise everything was peachy thus far

I decided to just go for a "money shot" - sarcastically referring to a "low blow" meant to end the tinkering session, and finally go to The Big Sleep for validation of Current Progress.
Whacked in

GDM Off(, 1T), tCWL 16->13, tCL 16->14, tCL 16->14, tRCDRD 16->14, tRP 16->14

wow what a surprise it doesn't POST loool

alright, one last one just for ****s and giggles;

tWTRL 12->10, tWR 12->10, tCL 16->14, tRCDRD 16->14, tRP 16->14

To my surprise, it POSTed!

it was even making progress in the dram calc bench without spazzing out!

You know, this was getting a little funny.

you know, I didn't plan for it to instantly succeed.
I thought these dual-rank thingies were supposed to have primary/trfc deficits compared to single-ranks, not out-perform them!

Now I just have various little subtimings to push, bah!

Wow 65ns latency just wow- I got my best to 66.8ns but not stable

 

[Bloax]

It's certainly very fortunate that within 24 hours of receiving a memory kit, I've already mostly tapped out the juice and established stability.

There seem to be some wacky performance regressions from ****ing with the subtimings vs. RDCalc thus far, and it's already getting a little hard to quantify the difference with this dinky-*** 3600.
So I think I'll just leave it "in the tank" so to speak until I get me a beefier CPU.

Hopefully, with a working IMC this time (and without any "commit suicide upon attempting to be returned" tricks)!

 

[Eduardo_Domingot]

I am trying to catch you but its been hard. I tightened things up alot though. Stress testing tonight.

 

[Bloax]

 

[Eduardo_Domingot]

Error city for me Bloax, I think I need to just face the music and buy faster ram. Might get a 4000mhz kit just to be future proof on a DDR4 platform long as compatibility is good.