Definitive guide to configuring the Ryzen 3900X/3950X and all other 3000 Series CPUs

[Michael Nager]

I'm not into pain anymore.

I just sold my i7-8700K system to my next-door neighbor.
That one was tweaked in many ways.
I de-lidded it and had very low temps with it. All cores were running at 4.972GHz.
I had SLI GTX-1080-FE GPUs in it and 32GB of DDR4-3200 MHz. RAM installed.
It seemed like I was always looking for a way to tweak it to get a faster benchmark score out of it.

Then, someone sold me an i9-9900k CPU and board.
I knew that it was a PITA to de-lid this one, so I got a good 280mm AIO for it.
I got 32GB of 4133MHz. RAM for it.
I got two 1TB Crucial Gumstick drives for it to run in a RAID configuration.
For DATA, there is a pair of 5TB Toshiba X-300 SATA drives in RAID configuration.

I put it all together in a high-airflow chassis with five 140mm Cougar Vortex fans. They move a shit-ton of air.
The SLI GTX-1080-FE GPUs moved over to this PC.
I went into the BIOS (ASUS ROG Strix Z-390-E Gaming) and set the RAID up. I set the Intel XMP (Extreme Memory Profile) to XMP-one.

I was ready,
Let the tweaking begin!
I started trying to OC the CPU.
No Joy.
no matter what I did to tweak the CPU, it failed to boot. After a few days of screwing around with it, I set the XMP-1 memory profile and left the CPU on stock settings.
It booted and ran fine. It also scored significantly higher in benches than the 8700K ever did.
I looked and found that it was boosting to over 5GHz. speeds. (not sure how many cores though)

Thinking about it, I decided to just leave it alone to do its own thing. It works, and I'm happy with it.

I'll probably set the Ryzen box up to do its own thing at first and see how it does.
I don't want to constantly massage my PS anymore. I want to just do some gaming with them.

The 3950X system I have is still what I consider to be my experimental system. I might go live with it and use it as my main system on my birthday next month.

Currently my main machine is an i7-4790K, which I have also delidded and used Thermal Grizzly Conductonaut liquid metal to replace the rubber that Intel laughingly calls TIM. It has been running with a mild overclock of 4.4 GHz on all cores since I first got it and in total the system has been up and running for more than 1,750 days 24/7 (4.8 years).

A couple of years ago there wasn't anything to spend my money on in the way of upgrading to a new system, so I decided to give the 4790K a new home.

Delidding made a lot of difference and here are the temps running AIDA Extreme on it:

The thing is that if you were sold an i9-9900K after Intel brought out the i9-9900KS then you're shit out of luck trying to OC it. The i9-9900KS was essentially Intel binning out the chips which would normally have been an i9-9900K that could be overclocked to 5 GHz (rumour has it that the "KS" in the CPU name stands for "Keep Spending").

If you do get a Ryzen system and keep your RAM then I would strongly suggest that you downclock the RAM to 3733 otherwise the Infinity Fabric will run at half of of what it should. At 3733 the Infinity Fabric would run at 1867 MHz; however above that the Infinity Fabric runs at Data Rate/2 which means that at 4133 your Infinity Fabric would run at 1033 MHz

@mtcn77
for anyone that put more than one machine together and can read a manual,
should not have any probs using LM.
even when i didnt care to do more than apply it (not removing it after couple of month and replacing it a few times to "saturate" block/HS),
twisting the block removed it without trouble from the cpu after being installed for +1y.
and using it for now +6y, showed me the "gap" between a quality (cooler) base and the cpu HS is a lot less
than what i would need from even the thinnest coat of ("normal") TP i could apply.

and it was easier to cover all of the HS evenly, when a (greasy/oily) TP was previously installed,
which always gives me a hard time applying new/different TP.


@Michael Nager
yet here i am, running aorus ultra with xmp loaded in bios, even managed to lower the timings from stock 18-22-22-42
(to 16-19-19-36/1, haven't tried more) at the same V, and on "crappy" micron E die.
no ram retraining happening.

Except of course that the i9-9900K is soldered.

That makes the prospect of delidding it not so attractive and as Der8auer has shown, the results even after delidding and lapping the CPU die are really not worth it.

I am running the Team Group Edition 3600 16-16-16-36 OC'd to 3733 with the same primary timings.

 

[moproblems99]

If you do get a Ryzen system and keep your RAM then I would strongly suggest that you downclock the RAM to 3733 otherwise the Infinity Fabric will run at half of of what it should. At 3733 the Infinity Fabric would run at 1867 MHz; however above that the Infinity Fabric runs at Data Rate/2 which means that at 4133 your Infinity Fabric would run at 1033 MHz

Not true. Many of us can and do run 3800 (or 1900 1:1). But it is not guaranteed.

 

[Zach_01]

Not true. Many of us can and do run 3800 (or 1900 1:1). But it is not guaranteed.

I can run 1900 1:1:1, but its not 100% stable. Its like 95+%...

 

[Michael Nager]

Not true. Many of us can and do run 3800 (or 1900 1:1). But it is not guaranteed.

You will only be able to run the Infinity Fabric at 1900 with a Ryzen CPU that is a silicon lottery winner.

It is VERY rare.

Even at that, running it and being able to run it stably are also two very different things.

I thought you were going to bring up that he could run the Infinity Fabric async with the RAM, but the big downside of that is the big latency penalty that incurs which is not worth it for 4133 RAM.

I can run 1900 1:1:1, but its not 100% stable. Its like 95+%...

I have never tried running the IF at 1900 and my RAM at 3800.

I am happy that it is rock solid at 3733 with 16-16-16-16-32 timings and a Command Rate of 1T with the IF running at 1867 1:1:1

I am pretty sure it wouldn't be stable and that the only way of even trying to make it stable would involve me having to loosen up the timings and I have no intention of going to CL18 and/or going to 2T.

 

[moproblems99]

It is VERY rare.

I wouldn't call it very rare. I put it somewhere between 40% to 60%. My 3700x will do nothing higher than 1600. A friend's 3900x will not go higher than 1600. Plenty of people here have no problem.

I am pretty sure it wouldn't be stable and that the only way of even trying to make it stable would involve me having to loosen up the timings and I have no intention of going to CL18 and/or going to 2T.

No idea what yours will or will not do but I didn't need to loosen timings or move to 1t.

Edit: I have some other ram with Samsung A-Die I need to try as well.

 

[RealNeil]

The thing is that if you were sold an i9-9900K after Intel brought out the i9-9900KS then you're shit out of luck trying to OC it. The i9-9900KS was essentially Intel binning out the chips which would normally have been an i9-9900K that could be overclocked to 5 GHz

My 9900K is one of the first batch ever sold. I bought it for jack-shit nothing from a friend that had it sitting on his shelf for months. After I bought it, it sat on my shelf for a lot longer.
I only recently built the system. Not trying to OC it anymore. It boosts to over 5GHz. on its own and that's that.

If you do get a Ryzen system and keep your RAM then I would strongly suggest that you downclock the RAM to 3733 otherwise the Infinity Fabric will run at half of of what it should. At 3733 the Infinity Fabric would run at 1867 MHz; however above that the Infinity Fabric runs at Data Rate/2 which means that at 4133 your Infinity Fabric would run at 1033 MHz

I already own the Ryzen 3800X CPU, an ASUS Crosshair VIII Hero Mainboard, and 32GB of GSKill RipJaws-V DDR4-3600MHz. RAM. (two 16GB sticks) I plan to run it at stock speeds.
So, according to your statements, my Infinity fabric should run at 1800MHz. But if it doesn't, I'll blow my nose with it.
I have two Vega-64 GPUs for this box, but I may sell them for a pair of 5700XT cards.

 

[pcwolf]

@Michael Nager I want to thank you for your musings in post #123 regarding the relationships between AGESA - XMP - SPD and DRAM training. For all the back-and-forth of this thread, that post gave me the most illuminating ideas to understand what the devil is going on between the motherboard and timings.

I have decided upon and been cursed with wrestling the ASRock X470 Taichi into submission. Not there yet, but your ideas and approach to things helps me in my understanding of the variety of tech factors pulling things back and forth. The Thaiphoon readouts of timings on my DRAM kit are also much lower than my successful experience so far in dialing them in with traditional overclocking procedures. So ... I appreciate your different approach, and your different experience as you relate it to us.

One of these days, I will go back to the start of this thread, follow your recommendations, and see what I will see.

 

[Michael Nager]

I already own the Ryzen 3800X CPU, an ASUS Crosshair VIII Hero Mainboard, and 32GB of GSKill RipJaws-V DDR4-3600MHz. RAM. (two 16GB sticks) I plan to run it at stock speeds.
So, according to your statements, my Infinity fabric should run at 1800MHz. But if it doesn't, I'll blow my nose with it.
I have two Vega-64 GPUs for this box, but I may sell them for a pair of 5700XT cards.

Up to 3600 if the FCLK is set to auto, then it should run at 1800.

It is only when you go above 3600 that you need to go in and set the FCLK manually.

 

[pcwolf]

I'm successful now at 3733Mhz 16-16-16-31 and 1867::1867 CR1
3800 is a No Go

 

[Michael Nager]

I'm successful now at 3733Mhz 16-16-16-31 and 1867::1867 CR1
3800 is a No Go

If that is stable then be very happy, you have a silicon lottery winner.

Here are the timings as far as I have tightened them.

I have posted two pictures, one is how the timings are applied, the other is the profile and the timings I had to leave on auto to preven Ryzen Master from wanting to reboot every time I made any kind of changes:

1) The timings applied to my RAM:

As you can see I am running the 3950X with SMT turned off, this not only gives me the superior compute power of Cores over Threads (the compute power of a thread is about 65% of a Core) but also to punt on 100 MHz per core above and beyond what is stable running the CPU with SMT On (4.3 GHz).

And now the timings as they have been applied to the AGESA portion of the BIOS:

Now of course YMMV but it might help.

If you had told me that you were wanting to tighten up your timings I would have posted this earlier as an aid for you and I apologise for not doing so.

I just noticed I had set the Peak Vcore voltage to 1.30626 (somebody had asked me about the different voltages and I went through them and when going back down I didn't go all the way back to 1.3 Volts, that's sorted now).

 

[pcwolf]

No apology needed. It is only through the generosity of yourself and other TPU posters that I have gotten as far as I have!

 

[Michael Nager]

I said that I would get back to you guys after I had installed the Der8auer kit described in the following video:

The temps are lower, it is not miraculous, but running CineBench it is now able to punt in 160 Watts at a lower temp, it has to be said though that this is with a FRESH application of Thermal Grizzly Kryonaut and my experience has been, that it takes a bit of time and the system running for a week or so for the temps to go down another two to three degrees.

So I am very happy with it.

It is a bit finicky to install, but it is pretty straightforward.

Is it worth it?

Yes.

 

[TheoneandonlyMrK]

I said that I would get back to you guys after I had installed the Der8auer kit described in the following video:

The temps are lower, it is not miraculous, but running CineBench it is now able to punt in 160 Watts at a lower temp, it has to be said though that this is with a FRESH application of Thermal Grizzly Kryonaut and my experience has been, that it takes a bit of time and the system running for a week or so for the temps to go down another two to three degrees.

So I am very happy with it.

It is a bit finicky to install, but it is pretty straightforward.

Is it worth it?

Yes.

I can't believe that's a thing in a way.

I mean what an opportunity for differentiation missed by aio Devs(this is my main point for clarity)

It's too dear and wouldn't work in any way for me, I am happy making my own mounts if required for cheaper.
But I went full coverage mobo block so no use here.

I would do something similar if it would help though.

 

[lorry]

I said that I would get back to you guys after I had installed the Der8auer kit described in the following video:

The temps are lower, it is not miraculous, but running CineBench it is now able to punt in 160 Watts at a lower temp, it has to be said though that this is with a FRESH application of Thermal Grizzly Kryonaut and my experience has been, that it takes a bit of time and the system running for a week or so for the temps to go down another two to three degrees.

So I am very happy with it.

It is a bit finicky to install, but it is pretty straightforward.

Is it worth it?

Yes.

You say that the temps Are lower, but not miraculous. But you do not go into specifics?
That sounds a bit odd mate, to me if I was saying that it worked, I would be saying that it has dropped my temps by xc degrees

 

[Michael Nager]

I can't believe that's a thing in a way.

I mean what an opportunity for differentiation missed by aio Devs(this is my main point for clarity)

It's too dear and wouldn't work in any way for me, I am happy making my own mounts if required for cheaper.
But I went full coverage mobo block so no use here.

I would do something similar if it would help though.

Given the problems with my back, I cannot guarantee that I will always be in a condition to do the maintenance on the loop, which is why I stick with an AIO.

I have had two spine operations and have spinal arthritis.

You say that the temps Are lower, but not miraculous. But you do not go into specifics?
That sounds a bit odd mate, to me if I was saying that it worked, I would be saying that it has dropped my temps by xc degrees

I did say that I had just applied a fresh application of TIM, even at that, the temps are around two degrees or so lower than the temps were previously where the TIM had been applied for enough time to have reached a homeostasis.

I have had a lot of experience with Thermal Grizzly Kryonaut and it takes about a week or so of "pumping" for the TIM to be at its maximum efficacy.

So the time to ask me about specific temps is in about a week or so time when that has happened.

This is why I have not given specifics as of this time.

If it goes the way it normally does, then the maximum temps should be around four and a half to five degrees Celsius lower than they were.

So there is nothing "a bit odd mate" about what I wrote.

I gave my initial impression, which is positive, and my final impression will have to wait for a week or so.

 

[TheoneandonlyMrK]

Given the problems with my back, I cannot guarantee that I will always be in a condition to do the maintenance on the loop, which is why I stick with an AIO.

I have had two spine operations and have spinal arthritis.


I did say that I had just applied a fresh application of TIM, even at that, the temps are around two degrees or so lower than the temps were previously where the TIM had been applied for enough time to have reached a homeostasis.

I have had a lot of experience with Thermal Grizzly Kryonaut and it takes about a week or so of "pumping" for the TIM to be at its maximum efficacy.

So the time to ask me about specific temps is in about a week or so time when that has happened.

This is why I have not given specifics as of this time.

If it goes the way it normally does, then the maximum temps should be around four and a half to five degrees Celsius lower than they were.

So there is nothing "a bit odd mate" about what I wrote.

I gave my initial impression, which is positive, and my final impression will have to wait for a week or so.

You have skills though bro, even when I directly note my actual point in my post you go tangential.

See my post again.

Wasn't against others doing, using it or you.

Pondering why aio Devs don't do this is primarily it.

 

[lorry]

Given the problems with my back, I cannot guarantee that I will always be in a condition to do the maintenance on the loop, which is why I stick with an AIO.

I have had two spine operations and have spinal arthritis.


I did say that I had just applied a fresh application of TIM, even at that, the temps are around two degrees or so lower than the temps were previously where the TIM had been applied for enough time to have reached a homeostasis.

I have had a lot of experience with Thermal Grizzly Kryonaut and it takes about a week or so of "pumping" for the TIM to be at its maximum efficacy.

So the time to ask me about specific temps is in about a week or so time when that has happened.

This is why I have not given specifics as of this time.

If it goes the way it normally does, then the maximum temps should be around four and a half to five degrees Celsius lower than they were.

So there is nothing "a bit odd mate" about what I wrote.

I gave my initial impression, which is positive, and my final impression will have to wait for a week or so.

Erm no you said

system running for a week or so for the temps to go down another two to three degrees.

not that they had gone down 2 degrees or so. two totally different things

 

[Michael Nager]

Erm no you said


not that they had gone down 2 degrees or so. two totally different things

If the temps with the fresh paste had been as high as they were with the bedded in paste then I would have been happy enough.

That the temps are a bit lower makes me happy, because I know from experience that those temps will go down a bit more.

The thing is though, that I won't be able to say anything definitive about temps for about another week.

I'm sorry about being unclear in the original post - that was my bad.

 

[lorry]

I'm surprised that it made that much of a difference, doesn't you block completely cover the IHS then?

 

[Zach_01]

I believe (not know for sure) that external coverage is not the only definitive aspect of blocks. Blocks have a center (or core) of cooling, and must be placed upon the hottest area of the chip. For most CPUs its right on the center but for ZEN2 its not.
Any block that is not Ryzen3000 specific designed, missing and not cooling the hottest spot/area of ZEN2 package as it should. Then you relay on IHS and TIM to spread heat sideways.

 

[lorry]

I believe (not know for sure) that external coverage is not the only definitive aspect of blocks. Blocks have a center (or core) of cooling, and must be placed upon the hottest area of the chip. For most CPUs its right on the center but for ZEN2 its not.
Any block that is not Ryzen3000 specific designed, missing and not cooling the hottest spot/area of ZEN2 package as it should. Then you relay on IHS and TIM to spread heat sideways.

yeah, I just don't remember what block he is using though

 

[Michael Nager]

I'm surprised that it made that much of a difference, doesn't you block completely cover the IHS then?

The problem is that all waterblocks for AIOs are designed for the hotspot to be in the middle of the PCB, and above all for there to only be one hotspot.

In the Ryzen 3rd Gen CPUs this design philosophy has been thrown out the window and the hotspot(s) are off centre, by a relatively large margin.

The business end of the cold-plate of a waterblock is where the heat exchanging fins are that the water passes through. For a single Chiplet CPU such as the 3600, 3600X, 3700X and 3800X less than 25% of that business end covers the hotspot which is the Chiplet.

For this reason, for the abovementioned CPUs you are better off getting a good air cooler (I have used the Noctua NH-U12A) to cool them because 50% of the cooling potential will be applied to that single hotspot.

For dual Chiplet CPUs such as the 3900X and the 3950X under load, an air cooler reaches its limits, as I know from experience.

For those same dual Chiplet CPUs the business end of the cold-plate of the waterblock of an AIO has less than 50% of its cooling capacity being applied to the hotspots and what is worse, the majority of the finned area of the cold-plate is essentially just twiddling its thumbs.

Another thing to consider is that for every 10°C of temperature, 4% more energy needs to be applied to the CPU to retain the performance.

For example, if, at 65 °C the CPU needs 150 Watts to achieve a performance level then, to achieve the same performance level at 75 °C you would need to apply 156 Watts, going from 75 °C to 85 °C would require 162.24 Watts to achieve the performance level you would have at 150 Watts at 65 °C.

Thus the worse the cooling solution is, the more the problem of heat, and heat generation becomes exacerbated and compounded.

What the kit from Der8auer does that it shifts the waterblock so that more of the fin area of the cold-plate can be brought to bear on the hotspots and can transfer more heat to the circulating liquid.

I am using the AlphaCool Eisbaer 360 AIO cooler.

 

[lorry]

The problem is that all waterblocks for AIOs are designed for the hotspot to be in the middle of the PCB, and above all for there to only be one hotspot.

In the Ryzen 3rd Gen CPUs this design philosophy has been thrown out the window and the hotspot(s) are off centre, by a relatively large margin.

The business end of the cold-plate of a waterblock is where the heat exchanging fins are that the water passes through. For a single Chiplet CPU such as the 3600, 3600X, 3700X and 3800X less than 25% of that business end covers the hotspot which is the Chiplet.

For this reason, for the abovementioned CPUs you are better off getting a good air cooler (I have used the Noctua NH-U12A) to cool them because 50% of the cooling potential will be applied to that single hotspot.

For dual Chiplet CPUs such as the 3900X and the 3950X under load, an air cooler reaches its limits, as I know from experience.

For those same dual Chiplet CPUs the business end of the cold-plate of the waterblock of an AIO has less than 50% of its cooling capacity being applied to the hotspots and what is worse, the majority of the finned area of the cold-plate is essentially just twiddling its thumbs.

Another thing to consider is that for every 10°C of temperature, 4% more energy needs to be applied to the CPU to retain the performance.

For example, if, at 65 °C the CPU needs 150 Watts to achieve a performance level then, to achieve the same performance level at 75 °C you would need to apply 156 Watts, going from 75 °C to 85 °C would require 162.24 Watts to achieve the performance level you would have at 150 Watts at 65 °C.

Thus the worse the cooling solution is, the more the problem of heat, and heat generation becomes exacerbated and compounded.

What the kit from Der8auer does that it shifts the waterblock so that more of the fin area of the cold-plate can be brought to bear on the hotspots and can transfer more heat to the circulating liquid.

I am using the AlphaCool Eisbaer 360 AIO cooler.

I have both (Alphacool eisbaer 360 and zde8sur) but yet to use either.
Eventually I will be usinthe AlphaCool r45 420 and the Optimus foundation block

 

[BrainMuncher]

Well you guys were right about the offset voltage being useless, what a shame, I thought that was such a good way of doing things. You could just reclaim back any extra headroom of your individual chip while keeping all other behaviors intact and without any risk. And if in any scenario you were power limited it would net you some extra performance from higher boosts.

So I tried out the whole PBO thing and that was also a bit pointless for me, seems like a feature for extreme overclockers. All it does is let the chip use more power but since the stock voltages are so astronomically high it doesn't help you because you end up thermally limited instead. And the chip just doesn't boost much higher that normal, I saw 4.225 for like one microsecond instead of the normal 4.2, not very useful, just extra heat, the opposite of what I want.

My mobo had two extra methods for setting the voltage under "AMD overclocking" - a second override, and a "max voltage offset".

Not exactly sure how this works but it does result in reduced performance. It seems to behave differently to the regular offset voltage setting, but I didn't bother to study it in depth since it doesn't do what I want. I'd also like to know what the difference between the regular voltage override and the "AMD overclocking" voltage override is. I can't really find any explanation.

What I would really want is some way to adjust the built in voltage/frequency curve, basically a working offset voltage setting. But it just refuses to deviate from the stock curve.

Anyway after a little experimentation I settled on a manual setting of 4.1GHz @ 1.25V. 1.225V was not stable. My goal was stock performance with lower volts/power/heat. Single core is actually the same as stock because even though it boosts to 4.2 at stock (and uses 1.475V to get there), I guess it doesn't stay there long enough to make a difference. All-core performance is slightly better than stock, at stock in sane all-core workloads like x264 it would sit at around 3950MHz at about 1.35V. But the temperatures are much lower, especially at idle and in light workloads, but all core temps are 7-10C lower as well. I was able to set everything in the BIOS without issue, although I did use Ryzen master to arrive at the values.

People say that the stock settings are pushing the silicon to it's limits out of the box, but it is clearly not. There is a huge difference between 1.475V and 1.25V. I'm guessing some of this stuff makes more sense with higher end models because you might actually need such voltages to get to 4.6GHz, but for the 3600 it's really over the top.

 

[moproblems99]

What I would really want is some way to adjust the built in voltage/frequency curve, basically a working offset voltage setting. But it just refuses to deviate from the stock curve.

Use Load line calibration. I set mine on 5 which reduces the voltage the most.