AMD Ryzen Threadripper 3rd Gen Overclocking Deep Dive, feat. ASUS ROG Zenith II Extreme

[1usmus]

1usmus, creator of the Ryzen Custom Power Plan, describes his experience overclocking AMD's new third-generation HEDT Threadripper platform, and shares a lot of information about cooling, CCX quality, per-CCX overclocking, memory tweaks, VRM options, and BIOS setup tips. Benchmarks are included, too.

Show full review

 

[cucker tarlson]

Fantastic read.

 

[kapone32]

Very interesting read. I actually have a question. I have a 2920X on the way. Will I be able to see an OC and performance gains like the 3960X due to the similarity in the way the CPU is wired?

 

[damric]

The apparent decrease in performance when raising base clock is probably due to the Windows 8/10 RTC (real time clock) which changes when you adjust base clock. It's not a real decrease however. Same effect as lowering your base clock in Windows 10 benchmarks to increase your apparent performance.

More info here:

(UPDATE 03/09) Windows 8 Results (Temporarily) Not Accepted At HWBOT - Benchmark Result Veracity Compromised

Read the full article @ HWBOT

hwbot.org

 

[Dixevil]

great read, thanks

 

[techguymaxc]

I was really looking forward to this processor release. I am not convinced it is the right fit for my workloads, however. It seems that most applications don't scale well beyond the 18-32 core HEDT CPUs that were available last year. Obviously this can change in the future, but I don't buy computational devices for future performance, I buy for what they can bring me today. Virtualization is also a viable option, though it does add cost due to the need for shared storage to support splitting my workflow over multiple VMs.

The way I see it there are 2 choices (for me):
1) buy the 3970x and get 99% of the performance of the 3990x in 95% of workloads.
2) wait for 4th generation and the (rumored) AVX performance boosts that it will bring.

It's a tough call, but thanks to AMD's execution of late I am inclined to wait for next gen.

 

[phanbuey]

I love the "the myth that it's hard to overclock is created by morons" line in the conclusion...

And then I read the process of actually having to overclock that thing and decided that I am firmly in the "Moron who think this is a huge pain in the a** to OC" category.

I must now learn to live with my new found limitations - great read & guide.

 

[W1zzard]

and decided that I am firmly in the "Moron who think this is a huge pain in the a** to OC" category.

we've all been there, everyone started from zero, don't worry about it

 

[HwGeek]

There is a mistake in R20 Single core graph- 2990WX listed with 529 pts(typo? 429 pts?).

 

[HammerON]

Excellent write up. Thank you for taking the time to do this

 

[1usmus]

Very interesting read. I actually have a question. I have a 2920X on the way. Will I be able to see an OC and performance gains like the 3960X due to the similarity in the way the CPU is wired?

you will get great performance because your processor does not use NUMA

 

[gamefoo21]

Ok let's start with this Asus marketing image.

First of all the signal from controller to doubler is single, but then it is split by the doubler and sent to each phase.

In the Asus system, the single phase consists of multiple stages, the signal from the controller is split from a single source and fed to each stage in the phase. There is NO balancing beyond the attempts to make each stage naturally balanced, but due to mass assembly and the desire for boards not to be $2000, the stages uses components that are +/- 5-20%.

How doublers work and can balance...

An intelligent controller can balance after dumb doublers. By reading how each phase is handling the load and actually telling the dumb controller to pull or push.

A smartish controller feeds a smart doubler the PWM signal. The smart doubler starts out just feeding both phases the same signal but monitoring both. Phase 1 is running a little harder than Phase 2, so the smart doubler actively shortens the PWM signal to one phase and can leave the other alone of lengthen it.

The Asus system... Well it's got stages so umm... Can't do that. It's faster because it's just running straight phases, the differences are incredibly small though.

How does load balancing work...

One doublers, you can actually shut down phases if you want.

On stages you are limited to actual phases. So you are often incredibly limited in having any phases powered down because the power specifications will often have minimum phases operating and wake up times.

Asus also tends to use older controllers that they rebadge and use for many years. So this has the advantage of being way cheaper, because you don't need fancy controllers and/or fancy doublers, so the savings can be passed on to someone... LoL

Marketing vs lying...

Asus tries very hard to using marketing spin to say parallel stages = phases. It's a straight up lie.

The truth...

1 phase and 2/3/4/5/or 6 parallel stage

= 1 phase

It's just that one phase is wider and has redundancy built in. Each stage will topple over and should naturally balance as loads increase. Though it does mean bad things can happen if a stage fails, all that load suddenly gets rammed through the other stage and the controller doesn't know. Which is another disadvantage, the controller only sees the phase, so as long as one stage is still functional it'll keep on keeping on.

Comments on the rest...

I really like the article and all of the in-depth information in it is the kind of stuff I love.

Also did EK provide you the half cover stuff about blocks? They did make an infamous TR1 block that was a massive failure because it was just an AM3 cold plate mounted under a TR1 mount...

There are lots of blocks and AIOs that do proper full cover Threadripper plates. That part felt kinda like it was trying to pot shot other sites and to push EK.

Just my opinion. I really enjoyed it overall but I'm sorry but your Asus board has an 8 phase VRM in reality. Please don't drink the Asus Koolaid it's bad for you.

 

[cellar door]

Damm dude... your articles are always soo damm good, thanks for another excellent article!

 

[rrrrex]

Aida64 memory tests are measured in GB/s, not MB.

 

[W1zzard]

Aida64 memory tests are measured in GB/s, not MB.

Fixed

 

[1usmus]

Ok let's start with this Asus marketing image.

First of all the signal from controller to doubler is single, but then it is split by the doubler and sent to each phase.

In the Asus system, the single phase consists of multiple stages, the signal from the controller is split from a single source and fed to each stage in the phase. There is NO balancing beyond the attempts to make each stage naturally balanced, but due to mass assembly and the desire for boards not to be $2000, the stages uses components that are +/- 5-20%.

How doublers work and can balance...

An intelligent controller can balance after dumb doublers. By reading how each phase is handling the load and actually telling the dumb controller to pull or push.

A smartish controller feeds a smart doubler the PWM signal. The smart doubler starts out just feeding both phases the same signal but monitoring both. Phase 1 is running a little harder than Phase 2, so the smart doubler actively shortens the PWM signal to one phase and can leave the other alone of lengthen it.

The Asus system... Well it's got stages so umm... Can't do that. It's faster because it's just running straight phases, the differences are incredibly small though.

How does load balancing work...

One doublers, you can actually shut down phases if you want.

On stages you are limited to actual phases. So you are often incredibly limited in having any phases powered down because the power specifications will often have minimum phases operating and wake up times.

Asus also tends to use older controllers that they rebadge and use for many years. So this has the advantage of being way cheaper, because you don't need fancy controllers and/or fancy doublers, so the savings can be passed on to someone... LoL

Marketing vs lying...

Asus tries very hard to using marketing spin to say parallel stages = phases. It's a straight up lie.

The truth...

1 phase and 2/3/4/5/or 6 parallel stage

= 1 phase

It's just that one phase is wider and has redundancy built in. Each stage will topple over and should naturally balance as loads increase. Though it does mean bad things can happen if a stage fails, all that load suddenly gets rammed through the other stage and the controller doesn't know. Which is another disadvantage, the controller only sees the phase, so as long as one stage is still functional it'll keep on keeping on.

Comments on the rest...

I really like the article and all of the in-depth information in it is the kind of stuff I love.

Also did EK provide you the half cover stuff about blocks? They did make an infamous TR1 block that was a massive failure because it was just an AM3 cold plate mounted under a TR1 mount...

There are lots of blocks and AIOs that do proper full cover Threadripper plates. That part felt kinda like it was trying to pot shot other sites and to push EK.

Just my opinion. I really enjoyed it overall but I'm sorry but your Asus board has an 8 phase VRM in reality. Please don't drink the Asus Koolaid it's bad for you.

You are a careful user, the advertising picture really does not show one element, namely the TPU, an intermediate element that "doubles" the signal, which then goes to two assemblies at once.
Yes, it cannot be considered a complete 16-phase solution, but it cannot be considered an 8-phase solution. That's why I put the Prime95 test results into the material. The 8-phase solution cannot provide a 55 degree temperature at 270-300A without active cooling or a very massive radiator.

 

[ypsylon]

In summation to this great line about "...the myth that it's hard to overclock is created by morons" . Despite ninja editing it shortly afterward it was very much there, but
I got to commenting on a free day.

Well on my side only a moron (in this instance person who wrote this line) overclocks a workstation CPU which can't OC past it boost clock anyway. TR3000 can game just fine at stock clock. You have to be clinically insane to consider 300W extra power draw sensible for permanent 4.3 GHz on all cores which you'll do anyway auto-boosting without overclocking... You think that 2-3 seconds in render time in Blender or 5 seconds in compile are worth recommendation at the cost of totally unstable system in the long run? Nobody doing serious job would ever consider that. OC is only for nerds with abstract synthetic benchmarks proving nothing. Out of the 63 graphs, if I calculated right, 25 are irrelevant synthetics. Game tests are complete garbage - show me anyone who owns TR 3960x/3970x or even older TR1xxx/2xxx and play games at 1080p. I know why you did that, not that dumb as author of this text implies, but... Nobody play games on TR @1080p, period. So basically you testing suite in 90% isn't worth squat.

Yes I was there at the beginning of OC, with Celeron 300A. One of the oldest rules in OC, if you can't OC CPU at least 10% then don't waste time. And you can't do that with any Ryzen. Auto All-core boost clock is maximum this architecture can do ( 3960x that's 4.3 GHz IIRC) because of high quality binning. Intel yeah kind of makes sense sometimes, AMD - don't make me laugh. And all of the results only confirm what I wrote above.


BTW: serious workstation note. Asus Zenith Extreme 2/Alpha have the worst memory compatibility of all TRX40 model for high density kits 128/256 GB. IMHO AMD should really limit TRX40 platform to 4 physical slots because TR gets really wonky with 8 sticks of RAM.

 

[W1zzard]

ninja editing

When phanbuey first posted, I checked the editing history, "moron" was never part of this article

 

[1usmus]

In summation to this great line about "...the myth that it's hard to overclock is created by morons" . Despite ninja editing it shortly afterward it was very much there, but
I got to commenting on a free day.

Well on my side only a moron (in this instance person who wrote this line) overclocks a workstation CPU which can't OC past it boost clock anyway. TR3000 can game just fine at stock clock. You have to be clinically insane to consider 300W extra power draw sensible for permanent 4.3 GHz on all cores which you'll do anyway auto-boosting without overclocking... You think that 2-3 seconds in render time in Blender or 5 seconds in compile are worth recommendation at the cost of totally unstable system in the long run? Nobody doing serious job would ever consider that. OC is only for nerds with abstract synthetic benchmarks proving nothing. Out of the 63 graphs, if I calculated right, 25 are irrelevant synthetics. Game tests are complete garbage - show me anyone who owns TR 3960x/3970x or even older TR1xxx/2xxx and play games at 1080p. I know why you did that, not that dumb as author of this text implies, but... Nobody play games on TR @1080p, period. So basically you testing suite in 90% isn't worth squat.

Yes I was there at the beginning of OC, with Celeron 300A. One of the oldest rules in OC, if you can't OC CPU at least 10% then don't waste time. And you can't do that with any Ryzen. Auto All-core boost clock is maximum this architecture can do ( 3960x that's 4.3 GHz IIRC) because of high quality binning. Intel yeah kind of makes sense sometimes, AMD - don't make me laugh. And all of the results only confirm what I wrote above.


BTW: serious workstation note. Asus Zenith Extreme 2/Alpha have the worst memory compatibility of all TRX40 model for high density kits 128/256 GB. IMHO AMD should really limit TRX40 platform to 4 physical slots because TR gets really wonky with 8 sticks of RAM.

Did you read the material?
You can get more productivity from the processor if you apply the right settings. Especially for you, I have described the intelligent overclocking of the processor.

In 1080p no one plays? There should be no video card tests in the material that concerns the processor performance.
I do not have three 2080ti to demonstrate the full potential of this processor in 4k resolution (I do not know what you dream).

 

[xkm1948]

Very cool. Would this overclocking (CCX or iOC) work in linux too? Most of us who get this for academia usage only use linux. Would be sweet if we can actually squeeze additional perf in linux as well

 

[phanbuey]

In summation to this great line about "...the myth that it's hard to overclock is created by morons" . Despite ninja editing it shortly afterward it was very much there, but
I got to commenting on a free day.

Well on my side only a moron (in this instance person who wrote this line) overclocks a workstation CPU which can't OC past it boost clock anyway. TR3000 can game just fine at stock clock. You have to be clinically insane to consider 300W extra power draw sensible for permanent 4.3 GHz on all cores which you'll do anyway auto-boosting without overclocking... You think that 2-3 seconds in render time in Blender or 5 seconds in compile are worth recommendation at the cost of totally unstable system in the long run? Nobody doing serious job would ever consider that. OC is only for nerds with abstract synthetic benchmarks proving nothing. Out of the 63 graphs, if I calculated right, 25 are irrelevant synthetics. Game tests are complete garbage - show me anyone who owns TR 3960x/3970x or even older TR1xxx/2xxx and play games at 1080p. I know why you did that, not that dumb as author of this text implies, but... Nobody play games on TR @1080p, period. So basically you testing suite in 90% isn't worth squat.

Yes I was there at the beginning of OC, with Celeron 300A. One of the oldest rules in OC, if you can't OC CPU at least 10% then don't waste time. And you can't do that with any Ryzen. Auto All-core boost clock is maximum this architecture can do ( 3960x that's 4.3 GHz IIRC) because of high quality binning. Intel yeah kind of makes sense sometimes, AMD - don't make me laugh. And all of the results only confirm what I wrote above.


BTW: serious workstation note. Asus Zenith Extreme 2/Alpha have the worst memory compatibility of all TRX40 model for high density kits 128/256 GB. IMHO AMD should really limit TRX40 platform to 4 physical slots because TR gets really wonky with 8 sticks of RAM.

When phanbuey first posted, I checked the editing history, "moron" was never part of this article

No that was me trying to be funny - I was paraphrasing. This is a really well written article -- one of those where it's worth reading through it more than once.

 

[Jism]

From 140FPS to 172FPS is defenitly worth it investigating in memory overclocking in any ryzen setup. +1 for this article.

 

[voltage]

come on Intel, get on with it and bring Tiger Lake already. I cant hold my breath much longer.

 

[TheGuruStud]

come on Intel, get on with it and bring Tiger Lake already. I cant hold my breath much longer.

I'll get you and your whole family a burial plot.

 

[Berfs1]

First of all, thank you very much for taking the time to dig deep into this. I wanted to say, Zen's safe voltage is 1.375V, Zen+ 1.35V, and Zen2 1.325V. You can go higher but you will face rapid degradation after those values without proper cooling. Also, considering it is ~20C to lose 100 MHz, I wouldn't consider 5 MHz/C a big difference. In most scenarios, the deltas will be 5-10C within coolers and ambients, so you will likely be looking at a variance of ~50 MHz.

In summation to this great line about "...the myth that it's hard to overclock is created by morons" . Despite ninja editing it shortly afterward it was very much there, but
I got to commenting on a free day.

Well on my side only a moron (in this instance person who wrote this line) overclocks a workstation CPU which can't OC past it boost clock anyway. TR3000 can game just fine at stock clock. You have to be clinically insane to consider 300W extra power draw sensible for permanent 4.3 GHz on all cores which you'll do anyway auto-boosting without overclocking... You think that 2-3 seconds in render time in Blender or 5 seconds in compile are worth recommendation at the cost of totally unstable system in the long run? Nobody doing serious job would ever consider that. OC is only for nerds with abstract synthetic benchmarks proving nothing. Out of the 63 graphs, if I calculated right, 25 are irrelevant synthetics. Game tests are complete garbage - show me anyone who owns TR 3960x/3970x or even older TR1xxx/2xxx and play games at 1080p. I know why you did that, not that dumb as author of this text implies, but... Nobody play games on TR @1080p, period. So basically you testing suite in 90% isn't worth squat.

Yes I was there at the beginning of OC, with Celeron 300A. One of the oldest rules in OC, if you can't OC CPU at least 10% then don't waste time. And you can't do that with any Ryzen. Auto All-core boost clock is maximum this architecture can do ( 3960x that's 4.3 GHz IIRC) because of high quality binning. Intel yeah kind of makes sense sometimes, AMD - don't make me laugh. And all of the results only confirm what I wrote above.


BTW: serious workstation note. Asus Zenith Extreme 2/Alpha have the worst memory compatibility of all TRX40 model for high density kits 128/256 GB. IMHO AMD should really limit TRX40 platform to 4 physical slots because TR gets really wonky with 8 sticks of RAM.

Well, workstation processors love higher frequencies when all cores are loaded. Since Zen2 does NOT have peak turbo for all core turbos, and manual overclock yields better results for all core, then it makes all the sense to do it. Also, if gaming is your focus, why the hell are you spending 1400$+ on a CPU for gaming? Don't even say "I want to use it for both WS and gaming", you can get a 3950X for that. And it isn't going to be 2-3 seconds or 5 seconds. Those fine tuned overclocks can help reduce on a several hour render even a few minutes. That is a few minutes less that they have to worry about. For big producers, time is money, and literally every minute counts. If I am not mistaken, if you use the boosting algorithm to match a manual all core overclock, you will actually be running a higher voltage which in turn means much higher power consumption. And that rule about 10% is bullshit. You can get MUCH BETTER power consumption when you tune your processor, regardless of if you go up or down in clocks. THAT is why we overclockers tune computers, because it makes the components more efficient and/or more performance efficiency (yes I said what I meant to say, that isn't a typo). And you are somewhat wrong about what the auto all core boost does. Where the boost of a 3900X would do ~4125 MHz with some offsets, I could take that same chip and do 4.3 GHz with manual tuning. Also, in regards to the memory compatibility, there is something really cool called QVLs. I wish people would read them for once before complaining "I CAN't DO 3200 MHZ IM GOING tO KiLL SOMEOnE".