After a long needed hibernation

[PiLsY]

TCKE 2 does indeed help - well found! This seems to be one of those hard to find borderline timings where you do several hundred percent in memtest then have a bunch of errors pop up within seconds. TRTP is another, I often find I can boot at TRTP 8, but at some point itll hard reset on me, then wont post until cmos is cleared. TRTP 12 fixed that (I see you already have it set chew*, just posting for anyone else who may read ).

You can try TRCDw 8 as well (assuming B-Die), though I dont think youll get that prime stable. My 3800x doesnt benefit much due to the single channel write speed, I suspect itll be harder to run but show more benefit on the 3900x (and 3950x).

[chew*]

Thx. Yah im usually good with timings. Studied there relationships with other timings for a long time. I just need nudging to not be lazy. Knowing 3800 nb was possible was the nudge I needed to not be lazy. Thx pilsy.

[PiLsY]

Beginning to wonder if i have a faulty X570i. With vdimm set to 1.4 in bios it was hitting 1.475v according to software in hci memtest last night. Runs great for hours then spews errors. Timings seem to make no difference, it just eventually craps out at anything above 3600mhz. Swap the same cpu and memory kit in the B450i and I get 3733mhz easy and 3800mhz with some tweaking, and all at 1.42vdimm and tighter timings as well. Its frustrating me badly.

[chew*]

Try dropping cke to 9 see if it makes a diff. In my notes gigabyte x370 disliked tight cke as well.

Ill plug x570 master in as soon as i get the mounts and sort out real world air clocks on taichi and these cpus.

[PiLsY]

Ill try that next, currently testing gear down enabled with 1T. MSI gets no benefit at all from that so I run straight 1T. Figured I may as well test if gigabyte have tuned for 1.5T...you never know

[PiLsY]

Interesting results...

3666mhz @ xmp 1T GDM disabled tried voltages all the way up to 1.47 vdimm, 1.12 vSOC, 1.05 vDDG. The furthest I got through HCI memtest was around 300%.
3733mhz @ xmp 1T GDM Enabled has run overnight to 650% or so at 1.28 vdimm, 1.05 vSOC and 0.985v vDDG. Going to leave it running while Im at work, but it looks and feels stable, much more so than it has in the past.

That compares to 3800mhz @ 16/16/16/16/32/48/288/1T @ 1.41 vdimm, 1.1 vSOC and 0.985v vDDG which passed 1000% HCI memtest on the B450i. Same memory sticks and cpu.

Massive difference between boards!

[PiLsY]

Reaches 1200% HCI...does stupid celebration dance then restarts pc. No post, cmos resets itself. Manually set exact same settings again...boots right up. Love it.

Decided the 1200% counts and moved up, 3800mhz now.

[PiLsY]

Try running Infinity Fabric above mem clock. Various configs between 3000mhz and 3466mhz with super tight timings and IF at 1900mhz are faster than any config I can run at 3600mhz+ for my sticks (B Die dual rank) and needs less SOC voltage.

Played a bit of PUBG last night after hitting 300% in HCI. Did about 2 hours gaming, then quit and sat in discord for a couple mins afterwards. Suddenly...screen went black. Turns out adding 10mv vddr seems to have fixed it. Went back to my 3733mhz profile and did a few restarts until I hit a post failure again. Kept adding vddr and doing loads of restarts until I got no more post failures, ended up adding 30mv. This is on an HCI 1200% passed profile. At this point I'm worried enough theres something funky with my boards vdimm power plane that Im going to email gigabyte support and ask if it should overvolt like it does.

Anyway, 3800mhz is HCI stable now but needs a lot of vdimm. 1933 Infinity Fabric ratio wont post so I cant really go higher. It feels like a bios limitation, but I can only do 101 bclk anyway due to nvme so thats my limit.

[chew*]

That may be because in the white papers on r1 showed that after 3200 memory it hit a hidden divider and we found bclk on a lower divider to exceed 3200 was much faster.

The problem was gen 1 and losing devices like on asrock we lost wifi. On cold the imc sucked so bad that I started pre tuning even lower and tighter so that when I went cold my tunes actually worked.



I was told that similar happens after 3600 memory divider which is why I was tuning for 3600 and not higher.

New agesa/bios released for taichi so scrapped all settings starting over fresh.

I'm on discord and apex a lot, feel free to find me under chewonth1s on origin chewonthis on steam

[PiLsY]

Good call, yeah theres dividers. Theyre pretty easy to find as it turns out...super pi will show a small performance drop after 2933, after 3200, after 3466 and after 3733 with all memory timings and fabric fixed. Im guessing theyre every 266mhz uclk throughout the range focused on JEDEC speeds, so at 1866, 2133, 2400, 2666, 2833 and so on, but I only tested the range above. My sticks arent optimal as I hit best latency at the mid point between dividers, I may have to pick up some 8gb sticks to mess with (I like the look of those C18 4500mhz 8pack/team group sticks). Waiting to hear back from a Gigabyte rep about the vddr overvolt still...

Forgot to add you on steam last night, will try and remember when i get home tonight

[chew*]

Noctua NH-D15 is teh best yo....

Well the best still is not good enough.

Doing decent but clocks are dropping to 4025 temps are 95c in 4k,8k 12k prime type loads @ stock....

Fyi its definitely not a fluke. The air coming out of the noctua is extremely warm in an A/C room around 70f

Charles you said 4.2 stable in what type of load on stock cooler?







Btw forgot to mention pilsy. I double checked my dual rank timings on x399. I was able to get really agressive with subs by using 2T CR and bank group swap enabled. That was with 4x16g b die c15 3466 bin.

[muzz]

Look at that beast pushing on the 5700XT!

[PiLsY]

Sorry Ive gone quiet...packed the board up today to send it back. Definitely have a vddr fault, it gets harder and harder to get it to post the higher I go over 1.42v. Its really hard to find stock as well, I have all the luck

[chew*]

I have been quiet due to testing for the 24/7 cruncher guys. Sysyems been running 4 1/2 days in prime on air. Clocks range from 4025-4175 load dependant in prime custom. 4k,8k,12k etc brings the lower clocks. Crunching is probably not as harsh so may see the higher end of clock speeds. Ill post it when its done.

Muzz yes its definitely close to the 5700xt but theres space for less than a pinky finger but its not touching. For card removal fans must be removed. Cant get at release. I prefer drp4 over this for set it and forget it and visual appeal and fitment. This however is better for testing as i have multiple board base mounts. I can do quick board swaps much faste since i have a second base mount already installed on another board.

[PiLsY]

It looks like the IO die counts towards the power budget when allowing the cpu to boost at stock or with PBO. My 3800x will actually boost itself beyond 4500mhz on multiple cores (2 cores average over 4500mhz during an R20 run) with PBO disabled when running really low memory speeds and voltages. Noticed while testing the vddr fault on my board, voltage config was 1.02v SOC, 0.87vDDG, 1.15vDDR, CPU volts at normal and settings were 2400mhz DDR / 1600mhz IF / CPU multi on Auto, PBO disabled.

Temps are lower too. I really didnt expect the IO die to make such a massive difference but it does. I wonder whether this would get overall better performance due to the cpu boosting higher or whether this is only really observable under really heavy multi threaded loads (in which case a rendering setup may perform better whereas a gaming setup running light multicore loads may perform worse). A 3800mhz memory setup costs around 180mhz average cpu speed if allowing stock boost behaviour as the IO die uses around 20% of your power budget on a 3800x when running prime95 (no AVX). Depending on the setups your cruncher guys run this may need testing. I havent done that for years, but back in my SETI days cpu speed was always king.

PBO acts differently when running high memory clocks too. If you disable PBO you will see higher average clocks on your best cores (as per Ryzen Master) but lower clocks on the rest. PBO enabled raises the lower clocked cores but significantly reduces clock averages on the best cores.

Could be per board/bios, test setup was the Gigabyte X570 ITX on beta bios F4L.

[muzz]

Tough break on the duff board chew, you have 2 others to beatup now though!
I got the DRP4 Saturday, I'm going to come out and say I should have ripped the board out of the case to install it ...what a PITA, but I got it in.....I did a couple of quick R20 runs with the case still open, and it went up to a bit over 70*, I still need to test it with the case closed, but I didn't close it until last night(busy, and played some SOTTR).
There is very little space between the HSF and the RX590, I don't believe a D14, never mind a 15(No Chance), would fit in there with the card in the upper PCIE slot, as it's possibly 1/8".
Very Quiet.

Speaking of TRTP- I'm not sure Dumos crazy 75K 4800 A64 score is really stable, but his TRTP was 6 on that run according to that memtester grab!

Now that I have the DRP4 installed, I can screw around with working on PiLsY's Manual clocking stuff, it's been running with PBO only so far.

[chew*]

I'm sure what dumo is showing is no where near realistic but that's extreme ocing.

this however is self explanatory.

[muzz]

I'm sure what dumo is showing is no where near realistic but that's extreme ocing.

this however is self explanatory.

I've been on the road for the last 2 days, so hopefully I get some time to mess around tomorrow/through the weekend.
Dumo was running 1.99v on his ram with a fan
It's fun to see what is capable, even if just a quick grab.
When you gonna start on the other 2 boards, and which 1 1st?

[bluestang]

You have to Disable Gear Down Mode to get 3600 C15? Or is that only for Ryzen 2? Just wondering if it's a RAM thing, a 2000 Series/X470 thing and if 3000/X570 fixed it or just the chip, etc?

[PiLsY]

Its a ryzen thing, applies to all generations. Note its only if you want to run flat timings with an odd cas latency (13-13-13-13, 15-15-15-15 etc). You dont need to run GDM enabled to use 15-16-16-16 for example.

Interesting development with my Gigabyte board, finally got a dialogue with Gigabyte support and they have confirmed that their board overvolts too, though not as badly as mine. The solution is to manually set vTT, once you have it set manually it automatically adjusts itself upwards as you increase vDDR. The issue seems to be that its auto value is 1/2 bios set value and not 1/2 measured value, so any voltage drift on vDDR puts vTT out of whack. The only annoyance is that the measured value is not 100% correct, its a little on the high side. If you have the same board as me (Aorus X570 ITX) I recommend manually setting the vTT to half your reported value in bios, this will get you close enough to run most overclocks stable. To get to the very limit you'll need to slowly tweak this down from 1/2 your bios reported value. It will only be slightly out, vTT is pretty sensitive though so if you're almost stable but not quite its worth getting it spot on to see if it stabilises you. I only need to get it perfect to run 1:1 3800mhz, otherwise 1/2 bios reported value is fine.

For reference my board overvolts 50mv idle and 65 to 75mv load for vDDR. Setting 1.45v or higher without manually matching vTT gives a no post and auto cmos clear at SPD/2133mhz memory straight from loading optimised defaults.

[bluestang]

Ok, thanks on that. Always forget about it and drives me crazy when trying to get my GSkill F4-3600C15D-16GTZ to run properly. Of course the XMP profile is borked on them, but manual config is touchy on these to get rated 3600C15 with my 2700X on X470 Taichi. Always seem to forget some setting somewhere lol

[monza1412]

Its a ryzen thing, applies to all generations. Note its only if you want to run flat timings with an odd cas latency (13-13-13-13, 15-15-15-15 etc). You dont need to run GDM enabled to use 15-16-16-16 for example.

Interesting development with my Gigabyte board, finally got a dialogue with Gigabyte support and they have confirmed that their board overvolts too, though not as badly as mine. The solution is to manually set vTT, once you have it set manually it automatically adjusts itself upwards as you increase vDDR. The issue seems to be that its auto value is 1/2 bios set value and not 1/2 measured value, so any voltage drift on vDDR puts vTT out of whack. The only annoyance is that the measured value is not 100% correct, its a little on the high side. If you have the same board as me (Aorus X570 ITX) I recommend manually setting the vTT to half your reported value in bios, this will get you close enough to run most overclocks stable. To get to the very limit you'll need to slowly tweak this down from 1/2 your bios reported value. It will only be slightly out, vTT is pretty sensitive though so if you're almost stable but not quite its worth getting it spot on to see if it stabilises you. I only need to get it perfect to run 1:1 3800mhz, otherwise 1/2 bios reported value is fine.

For reference my board overvolts 50mv idle and 65 to 75mv load for vDDR. Setting 1.45v or higher without manually matching vTT gives a no post and auto cmos clear at SPD/2133mhz memory straight from loading optimised defaults.

This is superb info mate, have that same board, also overvolts Vdimm. I did find some annoyance related to the VSOC, sometimes undervolting to 0.9v from 1.1v if set at Normal/Auto or undervolting with the offset mode. Problem is that VDDG remains in a fixed value of 0.95v.

[chew*]

Ok, thanks on that. Always forget about it and drives me crazy when trying to get my GSkill F4-3600C15D-16GTZ to run properly. Of course the XMP profile is borked on them, but manual config is touchy on these to get rated 3600C15 with my 2700X on X470 Taichi. Always seem to forget some setting somewhere lol

The only thing edited to run flat out xmp is trfc.

132/192/345. That was almost a plug and play prime 95 run.

[PiLsY]

This is superb info mate, have that same board, also overvolts Vdimm. I did find some annoyance related to the VSOC, sometimes undervolting to 0.9v from 1.1v if set at Normal/Auto or undervolting with the offset mode. Problem is that VDDG remains in a fixed value of 0.95v.

Glad that helped, its driven me bloody mad for weeks. Especially as Gigabyte take nearly a week to reply each time its the longest short conversation I've ever had!

To sort out your weird fixed voltages...as you increase memory multiplier the AMD AGESA code automatically changes some settings in the background, you'll find them set to "Manual" under the Settings > CBS section in bios. This is the part that controls AMDs stock boosting features. You can't change these back to Auto when over 3600mhz, itll just switch them back again when you restart, so you're forced to set your manual voltages here.

This is off the top of my head, but should get you pointed in the right direction. In the Tweaking menu I set the memory multiplier and subtimings, vDDR and vTT and vSOC to my values then change every other voltage in here to "Normal" (NOT Auto!!!). Now go exploring the bios settings page and youll find everything you need under the CBS and AMD overclocking menus. AMD Overclocking is supposed to be the master setting, but I set everything in CBS to match the AMD overclocking section anyway. The CBS menu is for controlling stuff while using PBO and staying within AMD spec limits so you cant go over 140A EDC for example. The AMD overclocking menu is where all the full overrides are. I'll screenshot everything in bios and post some sort of guide in the next few days (RL work permitting). This won't be feature complete by any stretch, I've been blue team for the last 15 years and I'm still learning Red stuff but i'll expand it as I go.

I do wonder if all the X570 Gigabyte boards overvolt vDDR like this. If they've taken the lazy route here maybe they have elsewhere. I did ask the question if they can update the Auto vTT algorithm to set 1/2 reported value instead of expected, but I wont find that out for another 3 to 5 days lol.

Also worth mentioning - "Last Known Good Settings" when recovering from a bad boot does not restore any setting behind an "Accept" wall but manually saved bios profiles do, so make sure you save a profile as you work and restore that profile if you get a cmos reset. Last known good is no good at all. This is the same on my MSI B450 and the Gigabyte X570, so its safe to assume this is an AMD thing in general.

[monza1412]

Glad that helped, its driven me bloody mad for weeks. Especially as Gigabyte take nearly a week to reply each time its the longest short conversation I've ever had!

To sort out your weird fixed voltages...as you increase memory multiplier the AMD AGESA code automatically changes some settings in the background, you'll find them set to "Manual" under the Settings > CBS section in bios. This is the part that controls AMDs stock boosting features. You can't change these back to Auto when over 3600mhz, itll just switch them back again when you restart, so you're forced to set your manual voltages here.

This is off the top of my head, but should get you pointed in the right direction. In the Tweaking menu I set the memory multiplier and subtimings, vDDR and vTT and vSOC to my values then change every other voltage in here to "Normal" (NOT Auto!!!). Now go exploring the bios settings page and youll find everything you need under the CBS and AMD overclocking menus. AMD Overclocking is supposed to be the master setting, but I set everything in CBS to match the AMD overclocking section anyway. The CBS menu is for controlling stuff while using PBO and staying within AMD spec limits so you cant go over 140A EDC for example. The AMD overclocking menu is where all the full overrides are. I'll screenshot everything in bios and post some sort of guide in the next few days (RL work permitting). This won't be feature complete by any stretch, I've been blue team for the last 15 years and I'm still learning Red stuff but i'll expand it as I go.

I do wonder if all the X570 Gigabyte boards overvolt vDDR like this. If they've taken the lazy route here maybe they have elsewhere. I did ask the question if they can update the Auto vTT algorithm to set 1/2 reported value instead of expected, but I wont find that out for another 3 to 5 days lol.

Also worth mentioning - "Last Known Good Settings" when recovering from a bad boot does not restore any setting behind an "Accept" wall but manually saved bios profiles do, so make sure you save a profile as you work and restore that profile if you get a cmos reset. Last known good is no good at all. This is the same on my MSI B450 and the Gigabyte X570, so its safe to assume this is an AMD thing in general.

Thanks for your input, looking forward for your guide.
About the vdimm issue, this was posted by the Gigabyte Rep in another forum regarding to the Master board, I also saw overvolts in the Elite..

RAM Voltage

I had R&D look into this - actually any software reading of a voltage (even bios) can be off from the actual voltage at the cpu. It's an approximation. So we did 3 tests. What does bios say, what do the voltage points on the x570 master say, and what can we read from the cpu socket.

Set ram to 1.35 (xmp or manual, same result)
Save, exit, reboot.
Bios said 1.38
Voltage points said 1.37
Cpu socket said 1.364

So there does seem to be a 9mV "overshoot" in this test but to be frank that's really not a big issue. I get that we all want to set x and see x, dead on, no variance, just rock steady exactly what we put... But that's not really how voltage works. The number we put in bios actually goes through a large equation that also affects other voltages. It's all interdependent.

You would see the same thing with vcore. Software, voltage readout points, and testing the cpu socket will all be slightly different and rarely exactly what you typed in bios.

This is also a function of our philosophy when it comes to bios, temperatures, etc. We show you the value without any trickery. The same way we put vrm temperature probes smack in the middle. We could offset the probe a few mm and lower the temp, but what's the point? To what end? The idea behind readouts is to give you the most accurate information so you can tweak, optimize, or just check how the system is running. We are doing the same with voltage. We could set an offset to "show" 1.35v, even if the cpu is getting 1.364, but we don't. We are looking into ways to display a closer approximation of the voltage, ie 1.364 just like the cpu socket, but we won't do this through a magic hidden offset.

If you are dead set on getting as close to exact as you set you could try going 0.01 lower than you want, but even that is an approximation and will vary based on board, Psu, and cpu. The only true way to dial in an exact value would be to probe the cpu socket and test various settings until you hit the exact voltage you want.

Tl;Dr it should be fine. If you want to see a lower voltage then lower the input (~10mV) and see if its stable.

P. S. Ryzen master appears to just be reading the register (ie setting) from bios. Not what the voltage sensor says. The sensor is doing its job, but impedance affects reading vs reality.

https://www.overclock.net/forum/11-a...l#post28078624

Don't know if the Auto vTT algorithm would be the same in those models since I didn't see many people there have tried beyond 3800mhz or if the vdimm overvoltage will be worse passing 1.45v. I will be looking forward for chew analisis with the giga board, maybe he can also replicate that issue.