Ryzen - Return of the Jedi

[chew*]

Nice. I have it sort of down to a science now....i can actually bin memory sticks for subtiming strength via a profile i made. Not all are created equal and higher bin does not mean better it seems.

[Formula350]

Have you by any chance been able to determine how G.Skill bin their TridentZ?
For example my 3200 15-15-15-35 kit, would they test it to be certain to do X Speed at X Timings and then the kit is de-rated by one speed level. Basically translating into my kit being capable of 3466 15-15-15-35, OR possibly 3200 13-14-14-34 (I don't seem to have come across CL13 kits but, nevertheless, 12 seemed like quite a leap from a CL15 rated kit).
lol I just realized I had a dream about RAM last night Can't quite recall specifics but I felt as though... Was basically that I managed to get it to run at like 2800 2-2-2-10 (which I believe 8 is the lowest primary timing they can run heh)


Also I was curious... You've said that you feel the Taichi has the best RAM MOSFET setup on all the boards you've had so far, and I was wondering what about it makes you say that? Is it the brand/model of components used, their quality, the number of phases, or what?
When looking at some photos I've noticed that it appears as though the majority of boards only have a 2 Power Phase design for DRAM, but the Titanium seems to have a 4 Phase setup. Now I know that those ones specifically are NIKOS brand MOSFETs/drivers (4xPK632BA, 3xPK616BA), and apparently they are not the best quality (outside my expertise), but I don't know what the other boards are using on their power delivery. Those are controlled by a ROHM Powervation PV3205 Digital Phase controller, which they've used on a few other boards (Z170A, Z270, Titaniums and lower models), but seems like if they are employing a 4 phase setup, it's doubled, as that can only feed 2 phases.





Basically am wondering where in your opinion what the Pros and Cons are, and what changes would've made it better...?

[chew*]

Cleaner signals. Ram does not really need the ooomph that taichi provided but quality parts can provide cleaner signals. Not stressing it can also provide cleaner signals.

Of course this is just a guess...without proper equipment there is no way to tell.

Would need an oscilloscope for concrete evidence which would be fun to use....but for my needs really would sit around alot.

If i decided to want to do real reviews maybe but I am here more for the hobby...once it becomes a job the enjoyment factor dies.

[Formula350]

Set the DRAM to 3333, upped the NB volts one notch to 0.985V (0.975V idle, 0.971V load), and DRAM to 1.375V. Didn't go so well.
Set LLC to Mode 8 on CPU, and on NB. Didn't go any better (imparts a fairly high droop, way more than Auto did)
Set LLC to Mode 3 on both. Went atrociously bad! Sure voltage is more stable under load, but every worker thread except #6 failed within a couple minutes, which is new...

Guess I'll be leaving it on Auto!

UPDATE: Played around with under-clocking/volting the CPU (DRAM still at 3333), since I could do all that from Windows w/o a reboot... Was just running the default Blend and started out at 22x 0.9V (reported as roughly .88V loaded), and after a few minutes Thread 2 failed. I dialed the voltage back a smidge to 0.8750 and no more failed for awhile. Dialed it back more to .850V (.812V operating, load) and all was well for a bit longer. Then I started raising the CPU Multi (all while Prime was running that same Blend, with Thread 2 still down). Eventually I got to around 2.7GHz before Thread 3 failed. Let things settle for a minute and raised it to 2.8GHz; BSoD.
Oh, this was with a CLDO_VDDP set to 0.885V as well (NB set to 0.975V).

Am hoping someone managed to develop a program that is able to adjust other voltages, too. This all FEELS somewhat heat/voltage related, but then again, maybe it's just false positives caused by Windows 10? Who knows. It's hard to tell with AMD's ways of exhibiting instability (gracefully, as nice as that is, compared to a BSoD), which was kinda nice to see a BSoD from the 'under' testing, as I knew it was due to too much speed with too little core voltage.

[Formula350]

Anyone happen to know which of these three are what I'd need to input into the BIOS?



The situation is as follows...

• I'm running Prime95 with the Max Temps stress test setting to test 3.7GHz set by Ryzen Master (note: just opening K17TK [latest beta v0.9.1.10] seemed to have confused Windows somehow, as it instantly behaved as if I had applied something and set multi to 35X with VID of 1.275V [as per AIDA's CPU Tweakings page], despite K17TK showing the VID was still at 1.350V)
• Initially everything was indeed appearing to be running in "OC Mode". AIDA said VID was 1.2750 which clearly was wrong. HWiNFO shows the VID as a static 1.300V. The VDDCR_CPU was at 1.337V (which I know is what mine displays in OC mode). The VCore I think was like 1.375V.
• My first voltage change (since I wasn't sure what exactly was going on and I knew that 1.3V would be sufficient for 3.7GHz) was changing the VID to 1.30000V, which dropped the HWiNFO VDDCR_CPU to ~1.28V and VCore to 1.320V. From there I adjusted it down in roughly 0.01V steps.
• The window on the upper left is K17TK, which I'm using the adjust the voltage since it has no overhead like Ryzen Master (and I wasn't working with benchmarks). Important to note, is that K17TK's applied VID IS in fact reflected by Ryzen Master, thus both display the 1.2750V as in the screenshot.
• The voltage it's operating at is stable, not jumping around any, at all VIDs I've used.
• Prime95 has been continually running all during my voltage level changes. I set the new voltage as soon as the last worker thread finishes Test 1, and I let it pass two tests (the Test 2 that it has started when I apply voltage, and then the following Test 1 with its "Self-test" pass).

I know the voltages are applying because they lower in HWiNFO, as well as the temps drop a couple C each volt change.

Currently it's passed about 6 more since I further lowered it to 1.26250V (1.194V VDDCR_CPU [1.189V VRM], 1.232V VCore), so I'm pretty happy with that as it lowered the core temp by 8 to 10C over the original voltage, and the TOTAL Power has dropped about 27W (Package dropped ~14W).

I'd just love to know what I'm really looking at, so I can apply this in the BIOS. Takes forever to reboot, enter, change, boot Windows, see what it's at, then repeat the cycle another X number of times just to get it to where it is now! lol

Update: I'll assume that this means it's a case of massive VDroop?

UPDATE 2: Alright, so I dunno what was up with the VID/Voltage reporting that day, but the only one that was correct seemed to be the VDDCR_CPU of ~1.19V. I set the BIOS voltage to 1.25V first and it was way off. I set it to 1.20V and it was basically perfect, though was strangely drooping just a bit too much (I presume because BIOS was set to Auto and 1.35V, but CPU only requested 1.2V). Nevertheless, the seemed to be the lower limit of the stable voltage for 3.7GHz anyhow, and that additional smidgeon of droop was enough to cause instability, which I'm trying to combat with LLC on the CPU. Mode 3 was actually worse, and so I decided to just try Mode 1 instead. Might be more than needed, but it keeps it at a perfect 1.2V unloaded AND loaded, which is impressive. I'll probably set it to Mode 2 later today.

[chew*]

Ok taichi 2.36 bios seems better for b die. I think 2.4 adds some Hynix or DR support but does some weird stuff.

aida copy seems low otherwise it seems normal.





2.36 3333 "compatibility" bios profile.
Rename to .BIN
3333 compatible.txt

[chew*]

annoying board + very annoying cpu.

times suffered.

Regardless...

sub 7 5 gig done.

[os2wiz]

Formula I am reporting a major bug in bios 1.70 for the MSI X370 Titanium. Not eveybody is affected, but the cpu multiplier, if changed from auto in order to overclock the cpu, will cause the system not to boot. Even if you change it from auto to 36 , which is stock for an 1800X, it will not boot. This has nothing to do with memory overclocking. Even if you attempt to boot with memory at 2100mhz ( I have ddr4 3200) if you change the cpu multiplier off auto, the machine will not boot. Had to reset cmos several times and play with settings before I was sure it is the 1.70 bios. It was reported by another user on MSI forums as well. When we both reverted to beta 1.74 bios the issue disappeared. This is MSI's first flawed official bios for the Titanium.

[msimax]

annoying board + very annoying cpu.

times suffered.

Regardless...

sub 7 5 gig done.

nice man

[chew*]

nice man

it's slow.

Just to show proper scaling and how much time goes into my findings.......

Projected proper efficiency is roughly 9.5 seconds shaved per 100 MHz at rather horrible memory speeds.

This means the above run was "slow" ( equivalent of a 4.9 run ) still even admittedly a poor result it was enough to hit sub 7........it is not hard.

4.0


4.6

[BeepBeep2]

annoying board + very annoying cpu.

times suffered.

Regardless...

sub 7 5 gig done.

Hey, still better than everyone else so far in 5G 32M.

I plan to pick up a board and 1600X / 1700X this week before I head to LN2 bench party. I noticed the new ASUS Strix X370 has an external clock gen, I may give that a try w/ FlareX and post some 4G air results soon.

[chew*]

Its not deneb or thuban or llano when cold...that is for sure.

My setup had a coldbug...sorted that...still had a cold slow bug.

Asus is slow in 32m...if i claim my times sucked...and i still passed everyone nor did i use my own ln2 prepped c6h.

I am by no means suggesting use taichi for ln2 either lol...not until i sort out if board or cpu caused my nightmare.

Zen is on "fastest" pi board. Mine are both in rma...no clue how it runs cold. I just know i managed to kill both on air.

Fwiw...ln2 mode on c6h jacks up pll and chipset voltage. Made no difference on taichi.

[BeepBeep2]

Its not deneb or thuban or llano when cold...that is for sure.

My setup had a coldbug...sorted that...still had a cold slow bug.

Asus is slow in 32m...if i claim my times sucked...and i still passed everyone nor did i use my own ln2 prepped c6h.

I am by no means suggesting use taichi for ln2 either lol...not until i sort out if board or cpu caused my nightmare.

Zen is on "fastest" pi board. Mine are both in rma...no clue how it runs cold. I just know i managed to kill both on air.

Fwiw...ln2 mode on c6h jacks up pll and chipset voltage. Made no difference on taichi.

I figure it'll be a headache going in blind, not sure if I'll have the time to pre-test too much except one air session. Been reading up on memclk holes and stuff.

Trying to stay away from ASRock after some of the crazy bios bugs buildzoid pointed out recently, but the Taichi seems the most solid from them.

I'm not *too* worried about 32M performance, mostly only Cinebench R11.5/15 runs and light 3D (5770 / 4890 / 4850). I'd rather bench any 3D on Kaby Lake but that is money / time I don't have right now.

Definitely taking my Deneb setup for throwbacks, I've got a couple that'll probably get most of my time - http://hwbot.org/image/946242.jpg (UD5) + http://hwbot.org/image/806332.jpg (C5F).

If the Strix X370 lives after LN2 session it'll probably go into my 24/7 rig and I'll do additional testing

[chew*]

No comment on buildzoid.

Extreme overclockers either work around issues or sit the bench. My 2c.

[ZenEffect]

I figure it'll be a headache going in blind, not sure if I'll have the time to pre-test too much except one air session. Been reading up on memclk holes and stuff.

Trying to stay away from ASRock after some of the crazy bios bugs buildzoid pointed out recently, but the Taichi seems the most solid from them.

I'm not *too* worried about 32M performance, mostly only Cinebench R11.5/15 runs and light 3D (5770 / 4890 / 4850). I'd rather bench any 3D on Kaby Lake but that is money / time I don't have right now.

Definitely taking my Deneb setup for throwbacks, I've got a couple that'll probably get most of my time - http://hwbot.org/image/946242.jpg (UD5) + http://hwbot.org/image/806332.jpg (C5F).

If the Strix X370 lives after LN2 session it'll probably go into my 24/7 rig and I'll do additional testing

good luck with that, just be aware that you will find yourself over 10 seconds slower at 4ghz right off the top and you wont be able to pull into the same efficiency. If you want the same efficiency as us, I would recommend doing what we did. Spend 200+ hours in 32m exclusively and on ryzen only. lots of re-learning to do. the efficiency isn't magic or dumb luck, its hard work.

[BeepBeep2]

No comment on buildzoid.

Extreme overclockers either work around issues or sit the bench. My 2c.

Fair.
I'm just going out of state w/ no internet until the day before. The Ryzen platform just needs to be working, I'm not expecting great results but a feeler session on the platform.

good luck with that, just be aware that you will find yourself over 10 seconds slower at 4ghz right off the top and you wont be able to pull into the same efficiency. If you want the same efficiency as us, I would recommend doing what we did. Spend 200+ hours in 32m exclusively and on ryzen only. lots of re-learning to do.

Fair, but I said not worried about 32M or 32M efficiency because I don't have the time to prep and test.
I'm only concerned with having a working platform for the bench party.
If I don't kill anything, the Strix board is going in my 24/7 setup where I still won't care about 32M efficiency and later will purchase another board.

the efficiency isn't magic or dumb luck, its hard work.

Not sure what you're getting at, not expecting a free lunch. I'm just out of state from 7/7-7/13, the Strix board availability is beginning of next week, and the bench party starts 7/14.
The fact that Ryzen won't run on WinXP is already enough for me to dismiss 32M before the bench party because I don't have the time even for much OS testing.

[drmrlordx]

Hey chew*, it seems like some of your photobucket links are dead. Not trying to nitpick, but now that I finally have some more time to play with RAM tweaking (maybe) I find that some of the pics from 2 weeks ago or so are gone.

[chew*]

Yah... I need to go back and remove all [img] tags or pay photobucket. They said i exceeded my bandwidth.


I may talk to Aaron and just use my ln2 cooling gallery again. Less headaches and bigger images.

[Formula350]

Why not just use XS for the images? That's what I use...

I mean it's sort of a hassle, but I can't imagine it's any worse than use a third party website.
Here's my method:
Upload image then "Insert in line"
Click "Preview" on your post.
Right-click the link to the image in the preview and copy image location link
Click Insert Image button and paste.


And that's it.

[drmrlordx]

Oh okay, photobucket. $40/month ha! Whatever.

Yeah Formula350, that's what I have learned to do. So long as XS continues to allow it.

Oh on a different subject, I wanted to run something past Ryzen owners here to see if you get the same results as I: I was messing with some of the y-cruncher binaries the other day, and I found that the SSE4.1 binary produced the largest power draw on my 1800x. That really confused me. The performance was meh, but the heat and power draw were insane.

My 1800x @ 4.0 GHz, 1.4v vcore and level 3 LLC was pulling 250W peak total power (CPU + SoC) according to HWiNFO64, and the tdie was around 75C peak. Madness!

In contrast, the Zen-specific (ADX) binary that loads up by default topped out at 220W peak power and never got tdie higher than around 70-71C. And it was faster.

[BeepBeep2]

Picked up 1600X, Crosshair VI Hero and cheap 16GB SK Hynix H5AN8G8NMFR-TFC rated 3000 CL15 at Microcenter. Roommate picked up same, but Giga AB350 Gaming 3. She's sitting at 2933 CL16 GDM ON no issues on bios F7, just needed slight nudge on SOC voltage to get it past POST. No use in tuning her board further, she just needed something stable.

Excited to start testing soon, even if 32M is going to be trash.

EDIT:
F*** yeah its trash.

8Gx2 SK Hynix single rank MFR , needed 1.63v for 3225 CL14-16-16-34-62, tRFC 470, geardown on. All BIOS configurable subtimings manually set, played with ProcODT between 48 and 60, didn't make much difference regardless of clock. 2980 12-15-15-28-58 tRFC 400 was ~5 sec slower and needed 1.78v and 12-14-14 only <2800. Doesn't boot above 3280 with CL16 so I tried to dial in CL14. First runs were in the low 9's
Was only able to get the sticks to boot 2666+ in slots 1 and 3, not 2 and 4 which are recommended in the manual. Geardown off limited frequency to 2700 no matter what other settings I chose. Couldn't run high BCLK, RAM frequency was gradually more limited the lower the divider.

Looks like B-die and another board are in the future, this is probably okay 24/7 ram @ 2933 CL14-16-16 but nothing good whatsoever.

On the Gigabyte Gaming 3, old AGESA would not boot these above 2400, that board is now also doing 2933 easy on AGESA 1.0.0.6

[chew*]

Told you its difficult

32m and fast runs on ryzen are probably the most challenging thing you can attempt on ryzen. Its why i have been focussing on it.

I like challenges.

Thats why i dove into the 2x16g DR issues and sorted them out way before agesa 1006.

[BeepBeep2]

Told you its difficult

32m and fast runs on ryzen are probably the most challenging thing you can attempt on ryzen. Its why i have been focussing on it.

I like challenges.

Thats why i dove into the 2x16g DR issues and sorted them out way before agesa 1006.

It is, but its also fun. This is more tinkering than I've ever done on another platform, even intel. I don't think this kit is going to do any better in 32M, just can't get tight enough. Seems samsung probably tolerates much lower secondaries as well. Found BCLK hole from 102-103 too, couldn't train DRAM at high speeds.

I'll have to order a kit give it another whirl soon. Microcenter just didn't have any better options for ram. Seems SK Hynix is an okay choice for 24/7 users with the new AGESA but its definitely not fast.
You did really well with the 2x16G.

I didn't realize it until now but the Crosshair board has AM3 mounting available and so does the new Strix board if that means much to anyone.

[chew*]

It means alot. Part of the reason i sold my prime. Saves me from gaff mounting phase and ln2 pot.

[chew*]

New AMD 32m PI WR.

Attachment 132387