[dk_mic]

Could you have a look at these timings, if there is room for improvement / some timings are too loose or tight?
I can't run higher IF with current BIOS, so I want to optimize at 1900 MHz.
5950x with 2 x Dual Rank 16 GB G.SKILL 3600 CL16 B-Die.
DIMM Voltage is 1.49 V in BIOS, but HWInfo reads 1.52 V. I have a fan keeping them cool enough.
AIDA latency is 54.8 (pretty clean, but not stipped down win10 bench installation)
Seems I can't lower tRCDRD at all (can boot, but plenty of errors in TestMem5).

 

[Audioboxer]

Could you have a look at these timings, if there is room for improvement / some timings are too loose or tight?
I can't run higher IF with current BIOS, so I want to optimize at 1900 MHz.
5950x with 2 x Dual Rank 16 GB G.SKILL 3600 CL16 B-Die.
DIMM Voltage is 1.49 V in BIOS, but HWInfo reads 1.52 V. I have a fan keeping them cool enough.
AIDA latency is 54.8 (pretty clean, but not stipped down win10 bench installation)
Seems I can't lower tRCDRD at all (can boot, but plenty of errors in TestMem5).
View attachment 2534009

Around 54ns is basically the floor for a 5950x at 3800/1900. You might get into the 53.x range, but not like 53.1~2. So the question more likely shifts to your read/write/copy and if you can push more out of them. 54.8 can be improved though, when I say around 54ns is the floor I'm meaning 53.8~54.2.

You've modelled your timings around GDM 1T, so I'm not too sure what to recommend you try as GDM can interfere with timings. Instead, just so you can see what I mean, try switching to 2T/GDM disabled and see if you can run a 25 cycle TM5 as is.

48 ProcODT IMO is much higher than necessary on a 5xxx processor. 36.9 is usually where DR with a 5950x will sit around. Also to fix your tRFC2/4, the formula for calculating them is tRFC2 = tRFC / 1.346 tRFC4 = tRFC2 / 1.625.

Anta advised tRFC itself is best as multiples of 16, your nearest options if you take that advice would be 256 or 272. I would try aiming for 256.

 

[VPII]

Glad to see you have made some headway. I have Windows 11 installed on a separate drive but I'm not paying too much attention to it until MS release patches that are shown to fix the outstanding issues with Ryzen CPUs.

Hi @mongoled just wanted to share I manage to get my entire system working with the -20 all core vcurve optimizer, results maybe not as great but how I have done it is by doing a complete fresh install of Windows 11. Yup maybe a bad idea with the issues with AMD Ryzen but it appears to be resolved. Now I have done this with a vcore offset in the bios of -0.0875vcore which was not possible before but it brought down temps under CB23 from 87c ccd temp to 80c which is all I wanted. Here is my result. Funny though before I did the vcore offset my memory latency was around 55 which is great but too great for my memory speed and timings I think.

 

[tcclaviger]

Whelp, decided to inflict mental pain upon myself and ordered a second 2x16 kit, so running 4x16 on the updated strix x570-E II.

After a round of massive OC success with the new board, catching and taking quite a few top spots for non-ln2 single core benchmarks, the bliss had to end sometime.

It's, going? I guess?!

3933/1966 C14 was fairly easy at 32gb.
3733/1866 won't post, even at 18s flat, 500 trfc, super loose seconds and thirds with 64gb.

Tried a massive assortment of voltage, RTT, CAD, etc values. To no avail.

Getting just this to post took about 4 hours. Have just started turning knobs on dialing the timings tighter, but I did verify 14-13-8-13-21-34 boots with trfc at 260/193/119. Tcwl 14

When people said going 32 to 64 was a big speed compromise, I somehow didn't fully internalize just how drastic it is!

The 48 GB setup, which is 2x16 GSkill 4000c16 and 2x8 Patriot viper 4000s was drastically easier to tune fast. Had 8gb stick in secondary slots, saw only a single memory divider decrease in MT/s and timings were almost identical. Probably going back to this config, the hit to speed for my IMC using 64 is just too much.

To boot this requires setups to be manually set to 0, auto won't post, but 56-0-0 will get into windows at 1t, but very unstable.

DrvStr pretty much have to be what they are.

RTT also have basically no wiggle, but this is what auto set. Tried other values, 7-3-3, 6-3-3, 7-off-1, 6-off-1, 7-3-1 works best.

TPHYRDL is 26 on all 4 DIMMS.

Board doesn't have a ram boot voltage option like crosshair. I suspect that's holding it back quite a bit. Also doesn't have Digi+VRM section for RAM like crosshair, also surely not helping since I can't bump the VRM frequency.

Temps are fairly cool, 19c idle, 21c TM5 (underwater).

The GVKA 4000-16 kit is great for 32, don't advise for 64 unless you want a rough time or have a CPU with a beast IMC.

 

[ManniX-ITA]

Whelp, decided to inflict mental pain upon myself and ordered a second 2x16 kit, so running 4x16 on the updated strix x570-E II.

Wow, you really did put yourself in a tight spot
Kudos for the bravery!

Board doesn't have a ram boot voltage option like crosshair. I suspect that's holding it back quite a bit. Also doesn't have Digi+VRM section for RAM like crosshair, also surely not helping since I can't bump the VRM frequency.

Wow, this is really astonishing. Did you try another BIOS release?
For sure it's not normal at all. I've never seen an ASUS board without Digi+VRM section in the last years.
Bought a much cheaper B550 TUF Gaming and it does have it.

When people said going 32 to 64 was a big speed compromise, I somehow didn't fully internalize just how drastic it is!

Yes it's drastic... did you try with GDM On? What about higher VSOC and/or higher/lower VDDP?

 

[mongoled]

@tcclaviger
So upping vSOC voltage did not help at all with posting @ 3733/1866 ?

Im sure you tried many combos, I would have set vSOC 1.2v, ClkDrvStr: 120 ohms, all other settings on AUTO to see if it posts, unsure what vDIMM you are using ...

If it did post would take note of all AUTO setitngs that are used and based on those AUTO settings would selectivly choose what to tweak ..

 

[mongoled]

Hi @mongoled just wanted to share I manage to get my entire system working with the -20 all core vcurve optimizer, results maybe not as great but how I have done it is by doing a complete fresh install of Windows 11. Yup maybe a bad idea with the issues with AMD Ryzen but it appears to be resolved. Now I have done this with a vcore offset in the bios of -0.0875vcore which was not possible before but it brought down temps under CB23 from 87c ccd temp to 80c which is all I wanted. Here is my result. Funny though before I did the vcore offset my memory latency was around 55 which is great but too great for my memory speed and timings I think.

View attachment 2534127

If the decrease is definately related to vCORE offset I would revist that and tweak it so you are not getting that increased latency as the increased latency would be a sign that you are loosing performance in other scenarios ...

 

[Taraquin]

Well, I got bored and it's Black Friday sales here and couldn't resist the urge..

I bought a 5900X and a ASUS B550-XE board.

Going to be re-using my B-Die I am running in my Intel 10900KF rig now. They will be replaced by a 2x8GB Kingston Beast 3600C17's I could get cheap and that set goes to a buddy who still rocks a 2500K lol.

My sticks are G.Skill Trident-Z Neo 3600C16 which I have always run at 4400 17-17-27-38-328-2T @ 1.50v on Z490. I'm assuming 4400 isn't very efficient on AMD due to not running 1:1 IF so what should I shoot for? 3800C14? Any tips where to start lol? I have to completely learn both RAM and CPU OC again as I have not used any Ryzen CPU at all let alone clocked one..

It's under a full custom loop, hwlabs GTX 420 + 240 push pull, EK D5 + res, EK Supremacy EVO block (maybe buying a new block for it tho).

And what TIM should I use on a mirror lapped block. Does LM work on AMD? I have loads of Coollaboratory Liquid Ultra here for Intel direct die stuff or should I just use PK-3 / MX-5?

The highest you can get your IF running without whea 19 is the best. On 5900X 3800\1900 is usually the max, but you can be lucky and manage anything between 3800 and 4200. 3800cl14 is probably the best you can do, getting it at 1T with tweaked subs, depending on binning 1.5-1.55V.

As for OC I strongly recommend curve optimizer as it gives you the best of both worlds with high single core boost, high allcore boost and low idle temp\consumption. Try -30 allcore as baseline, if unbootable, try -25 and so on and use core cycler to identify the cores that need lower neg value. Hopefully you will end up with some cores hitting -30, some around -20 and some maybe at 0. You could try setting +50-200 on PBO aswell, but it might require you to lower negative CO values which reduces allcore clockspeed so might not be worth it.

 

[Audioboxer]

The highest you can get your IF running without whea 19 is the best. On 5900X 3800\1900 is usually the max, but you can be lucky and manage anything between 3800 and 4200. 3800cl14 is probably the best you can do, getting it at 1T with tweaked subs, depending on binning 1.5-1.55V.

As for OC I strongly recommend curve optimizer as it gives you the best of both worlds with high single core boost, high allcore boost and low idle temp\consumption. Try -30 allcore as baseline, if unbootable, try -25 and so on and use core cycler to identify the cores that need lower neg value. Hopefully you will end up with some cores hitting -30, some around -20 and some maybe at 0. You could try setting +50-200 on PBO aswell, but it might require you to lower negative CO values which reduces allcore clockspeed so might not be worth it.

Just a reminder for anyone doing this, there is a very good chance all core -30 will boot, let alone 20~25, but you must do stability tests, preferably with OCCT or corecycler that can hit one core at a time.

Most of my cores are -25~30, but my very best core taps out at -7, and the 2nd best is at like -15.

I doubt many people are actually stable when they just run a -20 all core, let alone a -25 or -30. You need to use per core, put some time into it. Only running stability testing that hits all cores at once is not ideal either, as your all core boost will be lower than anything an individual core manages to hit on its own. I've seen quite a few people go with like all core -30, claim stable, play games seemingly fine and then complain weeks later about random reboots on desktop.

Ironically, idle is actually the end game stability test for most Ryzen chips

Take a look at this for example, while those boosts are soo small they're only for like half a second or something, if my best core, 2, isn't stable, one of these boosts could trigger a random reboot. It's normal for your best cores not to be able to achieve as big a negative value on the curve. More-so, it's actually expected they should be running a lower negative value than the weaker cores.

 

[tcclaviger]

CO fully tuned, ranges from -4 to -28, ycruncher stable using core cycler, sse + avx + avx2 stable in occt using physical+virtual cycling, extreme and varying load.

I concur that flat -30 or -20 across the board is almost certainly not stable. For those that it is stable for, the majority it's only stable when prochot is limiting. Cool that chip down and let it boost to what the voltage table allows for and it'll almost guaranteed over post into a black screen rest (so sub 50c at full load, not achievable by most setups).

I'm running a chiller so...I get much higher effective clocks than typical, especially when I drop water under 10c.

DoS tuned at 48/47 with 1.281 get (for longevity reasons, it's stable at 49.25/48.5 with 1.36 get).

With 48/47 dos + co/pbo I get these roughly and consistently.
31700/1695 r23
1870/19200 GB5
8070/73000 GB4
722/13800 cpuz

I did try a bunch of soc voltages, but didn't try 120 ohm, will give that a shot.

At 3933 it was decent ram efficiency, 61,3xx 53.2ns for flat 14s, at 3666 it's 56,2xx and 55.7ns, bleh.

Going to try for another day or two for 3733, but if it doesn't shake out, just going 48gb triple rank setup, I know it'll do 1933 with tight timings.

I do have a 1900fclk hole, sadly. To be honest, my 3900x dominates this 5950x regarding FCLK and IMC, such a shame to lose IMC lottery on the zen 3 chip, which has such good cores.

Fingers crossed for IMC update on vcache CPUs...

 

[tcclaviger]

Doing some telemetry trickery, got it to hold slightly higher at same temp, 5130-5150 effective clocks instead of 5110-5120:

The ram though, kills anything outside of static CPU benchmarks, things like Corona, Blender, any sort of gaming etc is all suffering. Lost 14fps in SOTR lowest 1080 and more proportionally in 3dmark tests :/ Thoroughly not worth it.

Also figured out an interesting relationship between EDC and boosting when cold, which allowed for this, but is otherwise totally irrelevant to any sort of normal ambient based cooling setup and more so one without DOS:

System manufacturer System Product Name - Geekbench

Benchmark results for a System manufacturer System Product Name with an AMD Ryzen 9 5950X processor.

browser.geekbench.com

These were essentially effective for this test, it held what I set it to hold (yes this was PBO).

Wow, you really did put yourself in a tight spot
Kudos for the bravery!



Wow, this is really astonishing. Did you try another BIOS release?
For sure it's not normal at all. I've never seen an ASUS board without Digi+VRM section in the last years.
Bought a much cheaper B550 TUF Gaming and it does have it.



Yes it's drastic... did you try with GDM On? What about higher VSOC and/or higher/lower VDDP?

It has Digi+ but it doesn't have the RAM section of Digi+ like the Crosshair boards do. Artificial segmentation to drive sales of the more expensive boards. Sadly, they don't properly support the higher tier boards as well (I have a C6E and C7H, both still have launch day bugs...).

 

[Audioboxer]

Doing some telemetry trickery, got it to hold slightly higher at same temp, 5130-5150 effective clocks instead of 5110-5120:
View attachment 2534136

The ram though, kills anything outside of static CPU benchmarks, things like Corona, Blender, any sort of gaming etc is all suffering. Lost 14fps in SOTR lowest 1080 and more proportionally in 3dmark tests :/ Thoroughly not worth it.

Also figured out an interesting relationship between EDC and boosting when cold, which allowed for this, but is otherwise totally irrelevant to any sort of normal ambient based cooling setup and more so one without DOS:

System manufacturer System Product Name - Geekbench

Benchmark results for a System manufacturer System Product Name with an AMD Ryzen 9 5950X processor.

browser.geekbench.com

These were essentially effective for this test, it held what I set it to hold (yes this was PBO).
View attachment 2534138



It has Digi+ but it doesn't have the RAM section of Digi+ like the Crosshair boards do. Artificial segmentation to drive sales of the more expensive boards. Sadly, they don't properly support the higher tier boards as well (I have a C6E and C7H, both still have launch day bugs...).

Excellent! What were you telemetry values to achieve this?

 

[ManniX-ITA]

Doing some telemetry trickery, got it to hold slightly higher at same temp, 5130-5150 effective clocks instead of 5110-5120:

Nice, at which temperature the CPU was while running?

 

[Mach3.2]

Just received the 2*16GB 3600MHz 14-15-15-35 kit, I still can't POST with tCL of 15 or faster unless I turn on XMP and leave everything on default. Only then I manage to POST and boot into windows at 3600MHz 14-15-15 1T GDM on. AIDA latency is around 59.5ns at those settings, which seem to be way too high.

I starting to think my IMC just doesn't like dual rank B die sticks. I haven't really banged into the wall this many times when I was tuning single rank sticks. 🤕

 

[Imprezzion]

The highest you can get your IF running without whea 19 is the best. On 5900X 3800\1900 is usually the max, but you can be lucky and manage anything between 3800 and 4200. 3800cl14 is probably the best you can do, getting it at 1T with tweaked subs, depending on binning 1.5-1.55V.

As for OC I strongly recommend curve optimizer as it gives you the best of both worlds with high single core boost, high allcore boost and low idle temp\consumption. Try -30 allcore as baseline, if unbootable, try -25 and so on and use core cycler to identify the cores that need lower neg value. Hopefully you will end up with some cores hitting -30, some around -20 and some maybe at 0. You could try setting +50-200 on PBO aswell, but it might require you to lower negative CO values which reduces allcore clockspeed so might not be worth it.

Thanks for the clear hints! I will probably just run XMP 3600 16-16-16-36 first just so I can focus fully on the CPU OC first.

Can you, or anyone else who wants to, give me a very short TL;DR version of what and how to stress test a 5900X with cycling cores and such when I go PBO2 + curve optimizer?

My CPU, board, spare Hynix 3600C17 RAM just in case it won't play nice with DR B-Die, and the TechN block is in the mail. Should be here tomorrow. It will take me a while to actually swap the build over as the rads and tubing are extremely tight in my case and I have no room to manipulate stuff around so I hope I can squeeze the board in there behind the tubing without having to pull the GPU out of the loop.

 

[Sam_Oslo]

Hey guys. I got a few questions for you RAM gurus.

I'm really happy with my current 2x8 3600Mhz 16 16 16 16 36 and considering to buy 2 more sticks. I found both 2x8 SR and 2x16 DR with the exact same 3600c16 timings and have these 3 setup options:

1- 2x8 SR + 2x16 DR for a total of 48GB
2- 4x8 SR for a total of 32GB
3- 2x16 DR for a total of 32GB (replacing the current sticks with 2x16)
And I prefer them in this order.

My motherboard (the Gigabyte Vision D) have proven to have a really good topology, and the 5700G did a good job for pushing the 2x8 sticks to 4600c18 without any problem. It didn't need any excessive voltages, but I'm not sure how it will cope with 4 sticks, especially with the option 1 for 48GB.

Which option do you guys suggest?

 

[ManniX-ITA]

Which option do you guys suggest?

You should read the posts in this page.
4 DIMMs is a challenge, if they are all SR.
They are probably not going at the same speed and timings that are working with only 2 DIMMs.
If they are 4 x DR it can be a nightmare.
Mixing SR and DR which needs very different bus and termination settings it's truly masochistic.
Depends on what's your favorite degree of pain; I'd recommend 2x16 DR.

 

[Sam_Oslo]

You should read the posts in this page.
4 DIMMs is a challenge, if they are all SR.
They are probably not going at the same speed and timings that are working with only 2 DIMMs.
If they are 4 x DR it can be a nightmare.
Mixing SR and DR which needs very different bus and termination settings it's truly masochistic.
Depends on what's your favorite degree of pain; I'd recommend 2x16 DR.

Thx for a good and clear warning about the nightmare . I had my fears but this sounds really harsh.

Is this nightmare the same on all motherboards? I ask to find out if it's due to the limitations of IMC on X-series CPUs?. Then my 5700G may have a better chance with 4 sticks?.
Anybody have experience with 4 sticks on G-series CPUs?

 

[ManniX-ITA]

I ask to find out if it's due to the limitations of IMC on X-series CPUs?. Then my 5700G may have a better chance with 4 sticks?.

It's probably much easier with an APU but the underlying issue remains.
Driving 4 sticks is more difficult, especially if they are DR.
Mixing SR and DR it's looking for troubles.
I've not seen a single successful mixing on DDR4 but maybe I've missed it.

 

[dk_mic]

Around 54ns is basically the floor for a 5950x at 3800/1900. You might get into the 53.x range, but not like 53.1~2. So the question more likely shifts to your read/write/copy and if you can push more out of them. 54.8 can be improved though, when I say around 54ns is the floor I'm meaning 53.8~54.2.

You've modelled your timings around GDM 1T, so I'm not too sure what to recommend you try as GDM can interfere with timings. Instead, just so you can see what I mean, try switching to 2T/GDM disabled and see if you can run a 25 cycle TM5 as is.

48 ProcODT IMO is much higher than necessary on a 5xxx processor. 36.9 is usually where DR with a 5950x will sit around. Also to fix your tRFC2/4, the formula for calculating them is tRFC2 = tRFC / 1.346 tRFC4 = tRFC2 / 1.625.

Anta advised tRFC itself is best as multiples of 16, your nearest options if you take that advice would be 256 or 272. I would try aiming for 256.

You are right, 2T GDM off is not stable on TM5. I have now tried quite a bit to get it stable at 2T while keeping the primary timings (and looking at error codes and this https://www.overclock.net/cdn-cgi/i...ale-down/https://www.overclock.net/attachments/tm5_1usmus_download-jpg.2479357/)
- but no success. Kind of frustrating with so many variables to play with.
AIDA benchmarks show no real differences, so I am not sure what is there to gain. I don't think the DIMMs can have a stable 3800-14 1T profile at all -if that should be my ultimative goal.

Thanks for the advice on tRFC -I didnt realize earlier that MSI bios even supports the individual settings.
I have played with lower ProcODT, but seems unstable. Are there any downsides to have it at 48? IIRC i had to increase it step by step when initially tuning the profile to get it to pass TM5, a long time ago.