Ryzen - Return of the Jedi

[drmrlordx]

3.8 GHz @ 1.3v is pretty good, especially with the stock cooler.

Lightman, can you do me a favor? Try booting with the stock DDR4-2133 settings, copying over all the timings/subtimings by hand, and then slowly run the RAM speed up to 3200 MHz. Set vDIMM as high as you're comfortable with, though I would recommend 1.4v vDIMM and anywhere from 1.1v to 1.175v SoC voltage. Let me know how far you can get with that thing.

I did this on my Taichi (actually used 1.42v vDIMM and 1.175 SoC voltage) and got up to DDR4-3466 . . . actually it was 14-14-14-28 + subtimings from DDR4-2133 default, but I digress.

[Lightman]

3.8 GHz @ 1.3v is pretty good, especially with the stock cooler.

Lightman, can you do me a favor? Try booting with the stock DDR4-2133 settings, copying over all the timings/subtimings by hand, and then slowly run the RAM speed up to 3200 MHz. Set vDIMM as high as you're comfortable with, though I would recommend 1.4v vDIMM and anywhere from 1.1v to 1.175v SoC voltage. Let me know how far you can get with that thing.

I did this on my Taichi (actually used 1.42v vDIMM and 1.175 SoC voltage) and got up to DDR4-3466 . . . actually it was 14-14-14-28 + subtimings from DDR4-2133 default, but I digress.

I copied stock timings from 2133MHz, adjusted CL to 14-14-14-34-48 1T from XMP and ran it all the way up to 3466MHz. Unfortunately it wasn't stable, even with 1.42VDIMM and 1.1875 VSOC. I even adjusted trfc to what chew* suggested - 256/416/560 but still got errors. Windows was bench stable, but running memtest would detect errors within minutes.
Next I dialled down MEM clock to 3333MHz and all seems to be good with trfc of 132/192/312. I lowered SOC voltage to 1.1V and DDR to 1.375V (first dab in the dark, not tested lower yet). MemTest completed 1 pass without errors which is good enough quick check.

Now onto what this 2h long exercise brings:

(I've used mBench cmd app to get these measurements)

R7 1700 3815MHz DDR4 3466 ManualTiminigs CL14-14-14-34-48
Access latency 65.7 ns

Read datarate (INT) 19948 Mb/s
Write datarate (INT) 11418 Mb/s

Read datarate (MMX) 22879 Mb/s
Write datarate (MMX) 12656 Mb/s

Read datarate (SSE) 27684 Mb/s
Write datarate (SSE) 50105 Mb/s


R7 1700 3815MHz DDR4 3333 ManualTiminigs CL14-14-14-34-48
Access latency 67.8 ns

Read datarate (INT) 19665 Mb/s
Write datarate (INT) 11128 Mb/s

Read datarate (MMX) 22332 Mb/s
Write datarate (MMX) 12146 Mb/s

Read datarate (SSE) 27209 Mb/s
Write datarate (SSE) 48098 Mb/s


R7 1700 3815MHz DDR4 3200 ManualTiminigs CL14-14-14-34-48
Access latency 69.1 ns

Read datarate (INT) 19064 Mb/s
Write datarate (INT) 10811 Mb/s

Read datarate (MMX) 21971 Mb/s
Write datarate (MMX) 11741 Mb/s

Read datarate (SSE) 26594 Mb/s
Write datarate (SSE) 46253 Mb/s


R7 1700 3815MHz DDR4 3200 CL14 DOCP
Access latency 69.8 ns

Read datarate (INT) 18167 Mb/s
Write datarate (INT) 9054 Mb/s

Read datarate (MMX) 20685 Mb/s
Write datarate (MMX) 9935 Mb/s

Read datarate (SSE) 24859 Mb/s
Write datarate (SSE) 43533 Mb/s


R7 1700 3815MHz DDR4 2666 ManualTiminigs CL14-14-14-34-48
Access latency 76.8 ns

Read datarate (INT) 17091 Mb/s
Write datarate (INT) 9201 Mb/s

Read datarate (MMX) 19624 Mb/s
Write datarate (MMX) 10007 Mb/s

Read datarate (SSE) 23601 Mb/s
Write datarate (SSE) 38263 Mb/s

[chew*]

Most important of all is besides you guys gaining some stability did using my borrow jedec spec sub timings and porting those vs xmp give either of you guys a boost in performance....

If it did plus stability

now we are getting somewhere.

Speed at the cost of performance = pointless

Also trfc is weird but so far i have figured out...trfc 3 can not be higher than 1+2...but that is all i have figured out concrete info wise...finding a magical ratio eludes me so far....but it can impact performance.

Also if settling for 3200 with realistic daily main timings/voltages.....you can try pulling in 132/192/312 to 132/192/270. Lmk how that works out performance/stability wise.

Also my little trick was for taichi as i felt its docp subs were rather loose having used other boards....prime pro may need a different approach.

[Formula350]

Well heres a tip..

Xmp is intel specific and "unsupported" by AMD officially.

Jedec is industry standard and supported by all and 2666 is the highest officially supported speed by amd.

Drop to 2666 @ auto on all but primary timings....look at the subs 2666 strap sets...start from there as a guide line.

As far as the bclk question previous guy asked...

There was a rookie "rumble" comp at hwbot. One part of it was highest ref clock...money prizes etc...if i narrowed it down to which board and exact settings to use...it might have well been me doing the result.
<snip>

While I did indeed know that about the XMP being Intel's profile, and thus only added to AMD systems through boardmaker's kindness, I hadn't considered using 2666 as a reference point. So I may indeed just have to do that Thanks.

As for the influencing competitions... That's the part about it all that I said I do understand people's reasoning behind withholding info/tips on. At the same time it's what I find frustrating because we went from being a niche group of people willing to share everything and turned into a cutthroat segmented/factioned submarket that's fueled by competitions with real money prizes. On the one hand, it's great that overclocking has gotten to where it has, where we have such an interest that there are competitions, because like most advancements in life there is a trickle-down to consumers like me. On the other hand... all of these nice shiny features now are generally left unexplained or with really vague descriptions. (Unfortunately that really hasn't changed much, as in the past those 'help' tips on settings were generally hard-to-understand or poorly translated Engrish anyways).

Like I remember with my Llano APU system, which the OC guide I read might have even been by you Chew lol, but there was lots of help you could find with those. Might've been because it was an APU and not as much of interest to extreme overclockers, but it was details like changing the SATA from AHCI to IDE to help attain higher BCLKs that I feel might not be a shared nugget of info with Ryzen. Granted, I have no idea if that would even help, given the years that have passed, as AMD could've came up with something to help prevent the need for that. Alas, I can't even test that since A) the Titanium lacks BCLK beyond 103MHz, B) the Titanium lacks the option to change to IDE mode. (It's in there, deep in a submenu, but there are SO damn many different SATA menus, all with similar options. So I don't know if changing one and leaving the rest on Auto would even result in the change being applied...)

New progress on 2 dimm overclocking. 3466mhz achieved at 3.875 GHZ . I am uploading Attachment 132343Attachment 132344Attachment 132345 validations. I have screen shots of the bios but they a re too large a file to upload the image. I just added AIDA64 memory report only.

Bios settings core voltage 1.3625 LLC profie 1 actual 1.408 v. , SOC voltage 1.152 actual CLDO_VDDP .99 v , dram voltage 1.44 v actual
memory timings 2T CR CL 16-16-16-16 42 tRC 72 BGS disabled ProcODT 63.3 OHMS. BIOS 1.73 beta MSI Titanium

Nice, dude!
"LLC profile 1" is CPU LLC I presume?

Also, for your BIOS screenshots (using F12), if you open them up with Image Viewer, then use Snipping Tool and click the dropdown on NEW to set to Rectangular Snip, if you select the whole BIOS view and save as a JPG, it'll only be 130KB However, I always save as PNG which they're ~330KB and upload just fine (as evidence by my previous posts).


As for my own tweaking, I tried out the SubTimings from these 3200 "Moderate" tightened timings, but primary at 14-14-14-35 1T and they caused SPI @ 32M to fail after the first pass. heh

(Settings: No CPU OC or voltages, CPU NB @ 1.05V (1.056V), DRAM @ 1.360V (1.376V) ODT 48Ohm (likely forgot to Auto it after coming off 3333), LLCs all Auto except "CPU NB Power Duty Control" which was set to "Current Balance")

Went back and changed JUST SubTimings to these in the below screenshot, leaving all the above settings the same, and now it passes SPi32M; however, does fail AIDA's RAM StressTest after only 25seconds (though, see headscratch after the pic) lol On the plus side, the Cache stress, where I usually have the problems, went on for >11mins before I stopped it. Also had absolutely no problems playing Minecraft all yesterday or the day before, which laugh all you want, but at the tighter timings above, it crashed fairly often :P (I own a server so I am playing MC a lot for that reason, mostly spent 'babysitting' the teenagers that play.)

(hah I just noticed that it converts the PNGs I upload; it's 102KB instead of the 322KB uploaded file)


*scratches head* And yet, I played around with K17TK, upped only the P0 FID to 36.5x and re-ran tests, came back and am running the RAM stress again, but this time it's ... 7m30s so far w/o a problem. Perhaps my chip/system just doesn't like CPB and setting the multi (which with this tool, SEEMS to disable CPB for me) disables some of the auto clock/voltage management, stabilizing things? I dunno... Now it's up to 9m45sec. I mean it's either that, or previously running the Cache test for 11mins had gotten things sufficiently heatsoaked, somehow causing the IMC to hiccup. Because I ran the RAM stress immediately after the Cache the first time, only difference being the fact the system was running totally default for PStates, etc. Either way. Cool in my book Stopping it at 12mins so I can go get the settings out of the BIOS.


Lastly, I typed this yesterday but didn't get around to submitting it, so as for the last post, I haven't had a chance to dive into that. However, I did run P0 up to 3.8GHz (default voltage, under load is usually 1.275V) and everything was solid as heck. Better than I've been, at least; ran Minecraft for a solid 10hrs w/o a single crash.

EDIT: Updated with image and LLC setting. This board is silly though... I actually had AXMP set on these Tridents, which beings nothing changes except primary timings, I had everything else set manually anyhow. Nevertheless, I disable AXMP, set 2133, left all my subtimings on accident and restarted... D2 code >_> *shrug* Silliness. Anyways, time to try the "JEDEC Trick" heh

[Lightman]

Thanks chew* for hints! I will experiment a lot more in due time, but I suffer from chronic lack of time to play properly with my PC lately ... It only took me 4 days to even start playing with memory timings and basic OC
Holidays would help here, but they are 2 months away!

I'm now very happy with DDR4 3333MHz CL14 1 day stable system. Next step for me is to play with main CL timings to see how good my RAM and CPU uncore are
BTW moving from Haswell 4.4GHz daily to so far 3.8GHz Ryzen feels like the right thing to do. Almost anything plays and runs as good or better already, considering I play with 144FPS cap to match my 144Hz screen and my W10 install is as old as Windows 10 itself!

[Formula350]

Tried 3333 with 2666 Sub & Turnaround timings, results in:
03 - System Agent initialization before Microcode loading.

Maybe I wasn't supposed to set Turnaround as well? I'll go drop those to Auto and try again heh

EDIT: Alright, booted after changing that.
(Primary is 18-18-18-38 2T)
Unfortunately, setting to 3466 and I get the usual: F9.
Set the tRFCs all to Auto... F9.
Set the ODT to 60Ohm... F9.
Set the DRAM voltage to 1.4V... F9.
Set the NB voltage to 1.2V... F9.
Set all the Drive Strengths to *shrug* I have no base, so middle value of 30Ohm... F9
Set the tCWL to 16, since I noticed it was 14... F9
Set all the Drive Strengths to 20Ohm (lowest)... F9
Set all the Drive Strengths to 120Ohm (highest)... F9 :|
Set them back to Auto; Set BankGroupSwap to Disabled... F9
Set BGS_Alt to Disabled... F9
Set all SubTimings to value +2... F9
*throws in the towel*


Still have no idea if setting the DRAM VREF Voltage to half that of DRAM Voltage will help, or if Auto does that on its own. OR if I should set it lower/higher...

EDIT2: PFF-hA :| Set it to 3333 and it won't even POST with things configured as they are. So I guess that's fairly telling of... well something. :P lol

Final EDIT: Loaded my tightened 3200, changed it to 3333, and much to my surprise it is in Windows with those timings! Maybe I should've tried a lower CPU-NB voltage instead of higher *shrug* Tests for another day. Going to see where this road leads me at 3333.

[chew*]

Formula i have multiple boards timing x @ 5 is not timing x @ 5 across multiple boards. It is due to some running tighter/looser.

I will start testing other boards soon and i rma'd both my gigas so have those back soon.

Still trying to figure out randomitis of taichi in 32m still.

[Formula350]

Formula i have multiple boards timing x @ 5 is not timing x @ 5 across multiple boards. It is due to some running tighter/looser.

Uuuh lol

I mean... I might have understood you?
If so, you're saying that, for example using arbitrary details, if you set your Taichi's tWRWR to 7, and I set my Titanium's tWRWR to 7, and we put them under an oscilloscope (or whatever), their actual 'timing' would not be the same?


OR, as this just came to me after posting... Are you saying that if everything is set to Auto, a kit of A-DATA rated for 3200 15-15-15-35 1T would generate different SubTimings than a kit of G.Skills rated for 3200 15-15-15-35 1T, despite having the same memory ICs?

[chew*]

Saying that while 7 may be 7 on both boards performance of one board may be tuned very tigh and or lack compatibility.

That said a loosely tuned board may be able to pull that timing down to 3-4...and the tightly tuned or lacking conpatibility will fail post just trying to pull same timing down to 6.

Taichi i have to pull certain timings in tighter than my K7...my K7 failed to post...my taichi got faster and can pull in even tighter still.

Basically it means you can not port timings from one vendor to another and just expect it to work. What works for one may not for another.

It gets even more confusing...some timing on taichi i can not pull in as tight as my K7...for example K7 twcl 10? Np ...

Taichi twcl 10 F9 .....

[Formula350]

Definitely interesting then.

Also, I always through tCWL had to be just below, or at, CAS? IOW if tCL is 16, you'll want tCWL at 14-16. heh On my Llano system (and even my PhenomII I think), any time I'd try and change tCWL to anything less than being 1 below tCL, things would break. Goes to show I really haven't a clue anymore, if I even did in the first place!

Anyways, running at 3.8GHz default voltage (using K17TK) and running RAM at 3333 14-14-14-36 1T 1.36V (1.376V actual) using my 3200 Moderate SubTimings on previous page. 28mins into the AIDA64 stress test running on Cache and System Memory (they create the highest load/power draw), and no problems yet. CPU Temp is 72 (so 52) and RAM is 31C. Best of all? CPU-NB is only at 1.05V ^_^ I did not like seeing them apply 1.15V so willy-nilly like it does, which unfortunately I think is an AGESA change, not a board vendor one. Though it might be. I think Beta 1.72 was applying slightly less NB voltage than what Beta 1.73 does.

Bench results are quite nice, running on 100% identical settings/timings except for RAM speed.
Memory Bandwidth jumped way more than I figured it would.
Read @ 3200 = 53686 MB/s
Read @ 3333 = 56088 MB/s
Write @ 3200 = 52574 MB/s
Write @ 3333 = 55299 MB/s
Copy @ 3200 = 47551 MB/s
Copy @ 3333 = 49113 MB/s
Also there was a rather large and unexpected jump in FP32 test, which makes me think that the increased CCX communication speed has helped.
FP32 at 3.8GHz/3200 = 8652 KRay/s
FP32 at 3.8GHz/3300 = 8815 KRay/s


Also, apparently I was way off on what CLDO_VDDP is for. I thought it was something CPU related, like for the cores and what not, but I guess it's the DDR4 PHY voltage. Sounds like I may have overlooked an option to try for getting moar ram speed!

[chew*]

Ddr phy voltage is a reference voltage like the southbridge ref volts. Both are 2.5v by default...rarely any gains to be had there.

[Formula350]

UPDATE: NEVERMIND! CLDO_VPP is more like the internal actual PHY voltage, and DRAM VDDP being the external reference voltage.
As per: https://community.amd.com/community/...lets-talk-dram

Which... damn! I wish I'd have been keeping up on the blog... They've provided all the juicy info I've been hoping to find!!



Oh right, duh, reason for coming here! Pfft lol I was going to pass along these two things...
First, a quote from The Stilt:

"tRC, tWR, tRDRDSCL, tWRWRSCL and tRFC are basically the only critical subtimings (for the time being).

Setting the SCL values to 2 CLKs basically makes no difference to the stability, but results in a nice performance boost.
Minimum tRC, tWR and tRFC depend on ICs and their quality.

tCWL adjustment is broken in AGESA 1.0.0.6 beta, but it makes pretty much no difference either."

Second, a zero-overhead (non system polling) program by The Stilt, to read Ryzen timings. Figure it'll work a bit better than MemTweak for screenies
http://www.overclock.net/t/1624603/r...#post_26137022

"RTC (Ryzen Timing Checker)

Password: "RyzenDRAM"

x86-64 only.
Supports Zeppelin & Raven based Ryzens.
Let me know if (when) you find bugs."

[chew*]

Twcl...is sort of broken...board dependant i think...twcl 10 on taichi fails post. On k7 it did not.

This is why i use pi...while those may be the most crucial....other timings have an impact that bandwidth tests can not see.

Any timing not listed above can be the difference between 300th and 200ths per second in loops in 32m...32m is a fine art...one must ignore total output and pay far more attention to time between loops.

Faster loops can be acchieved two ways....tighter sub or more stable sub etc looser.....

99% of time later after tuned i run aida and my bandwidth looks nasty good and more balanced as in copy is much higher

[Formula350]

Looks to have the same wonky behavior as CAS. Odds = no bueno. I had mine set to 13, but BIOS read 14. I set it to 12 and BIOS reads 12 now.

Just booted with the "SCL @ 2". Hopefully the tCWL change doesn't impact my stability, or performance too greatly, so that I can get a good comparison on what the SCL change brings. Before and After...

====================== BEFORE =========================================== AFTER ======================


If only cache results were consistent, then I'd know whether or not it's a normal AIDA anomaly, or a result of tweaking :P
Nevertheless, sizable gains! More than I've experienced from general tweaking of primary timings! For example, 3200 at 22-22-22-52 vs 15-15-15-35


EDIT: I stopped paying attention to Pi results on my system when they started getting slower and slower as everything else's results began to get faster and faster! For example, my best 1m run comes in at 9.861s which was running at stock CPU under a Balanced plan with RAM @ 3200 and probably the release BIOS. Where as I run it recently and it is ~11sec. Ran it now while CPU @ 3.8 and it proves my point: 10.989s.

[chew*]

pi 1m ehhh not really a timing bench.....

PI 32m......absolutely a timing bench.

Fine tuning loops at its best

[Formula350]

The problem scales for me, though. 1M is just my way to judge if things are the same, or if they've gone back to normal. Granted, I'm in Windows 10, but on the same system with nothing changing except BIOS and Speeds (no patching, driver updates, etc), I don't see why performance has tanked in SuperPi.

Do you have a rough idea of what times, in Win10, a 32M run should pull off with CPU @ 3.8GHz and RAM @ 3333 14-14-14-35 1T (with the aformentioned SubTimings and the SCL trick)?

EDIT: Just thought of this... Wonder if the reason SPi tanking so badly is in any way related to why Task Manager is unable to figure out my CPU speed after I use K17TK to adjust P0? (literally my steps are: boot system, let it rest for 15-20min, load up K17TK, set FID to 38.0, apply, load SuperPi)
CPU-Z reads it correctly, but Task Man seems to think 38.0 multi = 3.3GHz (power plan is set to 100% Min/Max, so it generally reads 3.9GHz due to Boost)

EDIT2: At 37.75x I just posted a time of 8m 59.619s. Iinitial was 6s flat, Loop 1: 24.533, Loop 2: 45.644.
Running it again @ 38.25x. Initial: 6.813s, Loop 1: 26.205, Loop 2: 48.269. And just keeps getting worse.
LITERALLY all I did in between runs was use SnippingTool to take a quick screenshot of the times to compare, changed the FID (K17TK was still running), and started the run. I never really understood why Pi runs have such a huge margin of error. Doesn't make running it very fun when it's so damned inconsistent lol I'm afraid if I fart, it's going to tack on another 30seconds!
Aaand 9m 7.526s on this run, at 50MHz faster :|

[chew*]

do not fart it makes times slower lol

on avg @ 4.0 8:20 is realistic in win7 pc 3200 c12......

I think we calculated about 12 secs per 100 MHz and about 3 secs per bump in strap 3200-3333 and c14 to c12 was about 2 secs......

3.8 should be able to pull off 8:46 @ 3200 c14 and 8:43 @ 3333 c14 and since we have tuned w7, win 10 should be about same time.....W7 is slower no tuning.

I have pulled off a 8:26 on msi @ c12 3333 4.0 but....it should have been in the 8:13-8:16 range imo......being that it was 10 secs slow unless new bios pulled miracles......8:56 @ 3200 c14 3.8 cpu

Key is to watch the time between loops you will notice like 6-8-10 are slower than say 5-7-9 etc but the goal is lower the time between loops = progress

After tuning I ran aida I was like 57k read 56k write 55k copy.

very very balanced performance

[Formula350]

Played an old Win95 game for a few hours, killed a couple services, ran it at 3.825GHz and got the ol' NOT EXACT IN ROUND. Dropped it back to 3.80GHz and ran it again. Got my best ever, actually.

I don't quite understand Loop 23 and 24 though. Everything was predictable at 21sec per loop up until them, which ended up being ~19.5s/loop

At any rate. Good enough for this idiot! (I've literally done nothing in between today's earlier runs that yielded the 8m59s and 9m07s, and this one, other than gaming lol)


(heh you're running it with LAA turned on, I take it? )


EDIT:Eeyyy! Pre-sleep gift from MSI! Beta 1.74 was uploaded

[Dante80]

A quick question. I am using the Taichi board with G.SKILL Flare X Series 16GB (2 x 8GB) running at 3200Mhz CL14.

With the latest bios update (2.40), is it possible to buy an identical set and move to 32GB at the same speed/timings?

Lastly, a question. Here is the current memory.



If I remember correctly, I am running this on the XMP profile.



While in the SPD tab the profile tRC is given as 48 clocks, in the memory tab it shows that I have 73. Is that normal or do I have a looser timing somewhere?

[chew*]

You have to manually set it. Imo 52 is a happier setting than 48.

[chew*]

Played an old Win95 game for a few hours, killed a couple services, ran it at 3.825GHz and got the ol' NOT EXACT IN ROUND. Dropped it back to 3.80GHz and ran it again. Got my best ever, actually.

I don't quite understand Loop 23 and 24 though. Everything was predictable at 21sec per loop up until them, which ended up being ~19.5s/loop

At any rate. Good enough for this idiot! (I've literally done nothing in between today's earlier runs that yielded the 8m59s and 9m07s, and this one, other than gaming lol)


(heh you're running it with LAA turned on, I take it? )


EDIT:Eeyyy! Pre-sleep gift from MSI! Beta 1.74 was uploaded

23-24 is usually a nice jump on a good run.

Btw meant to mention to you...instead of clocking up in windows...try boot 3.8 then run everest. Looks like you got hit by rtc in your aida. I got hit on my b350 board clocking in windows...inflated my scores alot.

Thats why most of the time you see me testing in win 7. Not effected by those bugs.

[Dante80]

You have to manually set it. Imo 52 is a happier setting than 48.

So, the idea is to add tRAS + tRP, and then offer a small margin for stability, right? 48 is the direct sum of those, so the reason for 52 is the margin?

EDIT: Ok, flashed the new BIOS (old one didn't let you to move the tRC). Here is the result.



For reference, the voltage in the BIOS is set to 1.35V. Also, no idea whether that Trfc is normal/expected/ok or not.

[chew*]

That is what i found to be default for flare x. Trindent z in same SPD is 312. Trident Z in 3600 SPD was like 6xx.

Also i might add i previously tested 32g in taichi..i am retesting now.

DR 16g worked but last time 4x8 did not.

1.35v is like 1.37 real on taichi.

I am not sure about margins but we have noted performance loss if trc is to tight even when we are running 12-11-11-21.

32...is not optimal in fact it will not even boot.

B die seems to like 46-48 as the tightest trc @ lower speeds. Increase as you increase speed. Also seems to like even not odd numbers.

[Dante80]

Many thanks. If you do test, please let us now, my cursor is hovering over the buy button to get another 16GB.

Before I flashed 2.40 (1.0.0.6), I was using 2.20 on the MOBO. The reason I had Trc @73 clocks was simply because the BIOS did not have an option to tweak it. Now with 2.40 it seems to have one for all memory settings.

[chew*]

Here is a Taichi 32g baseline with some PERFORMANCE minded timings set....

May be able to extract more performance but first a baseline of stability is needed.......hope this helps.