• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

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

Joined
Feb 10, 2020
Messages
62 (0.13/day)
Thanks @1usmus for your amazing work.

I've been reading this page for a very very Long time but never really and (finally?) decided to Register just to thank you for your amazing and detailed work. I loved your last few articles (and especially your hard work and also very in-Depth Analysis of AMD+RAM article -- which is 2nd only to this one) because you've outdone yourself again - which is not an easy feat.

I wish I had the possiblities you have but in Absence of this I applaude you for pushing the Silicon (and yourself) further and further ((the possibilites are given to a lot of People, but most shy away from the time commitement it takes to truly live up to what they could acchieve or do with the Silicon in their Hands -- it's not "sexy", doesn't leave a lot of time for other Hobbies or gaming and thus most have a "done enough for this" menthality which is sad but makes you stand out even more :) ))




Thanks again for a very interesting read on my commute to work and some great Pictures (also not a given for every article, as it's "trivial" technically but still takes time and commitement to not just use pre-given Pictures or "snapshots while I'm at it" but use a neutral Background and highlight one piece at a time etc etc). Keep up the great work! :)
 
Last edited by a moderator:

1usmus

AMD Memory Guru
Joined
Jul 28, 2018
Messages
61 (0.06/day)
Location
Ukraine
Thanks @1usmus for your amazing work.

I've been reading this page for a very very Long time but never really and (finally?) decided to Register just to thank you for your amazing and detailed work. I loved your last few articles (and especially your hard work and also very in-Depth Analysis of AMD+RAM article -- which is 2nd only to this one) because you've outdone yourself again - which is not an easy feat.

I wish I had the possiblities you have but in Absence of this I applaude you for pushing the Silicon (and yourself) further and further ((the possibilites are given to a lot of People, but most shy away from the time commitement it takes to truly live up to what they could acchieve or do with the Silicon in their Hands -- it's not "sexy", doesn't leave a lot of time for other Hobbies or gaming and thus most have a "done enough for this" menthality which is sad but makes you stand out even more :) ))

Thanks again for a very interesting read on my commute to work and some great Pictures (also not a given for every article, as it's "trivial" technically but still takes time and commitement to not just use pre-given Pictures or "snapshots while I'm at it" but use a neutral Background and highlight one piece at a time etc etc). Keep up the great work! :)

Thank you very much for the comment :)
I will try to make it a rule to publish more often, at least once a month or two.
 

phill

Moderator
Staff member
Joined
Jun 8, 2011
Messages
9,999 (2.75/day)
Location
Somerset, UK
System Name Not so complete or overkill
Processor Ryzen 3900X
Motherboard Asus Crosshair 8 Hero (Non Wi-Fi)
Cooling Air :(
Memory Patriot 3200MHz 64GB
Video Card(s) 2 x EVGA GTX 1080 Ti SC Black Edition
Storage 1 250GB Crucial SSD for OS/Programs, 2 500GB Crucial SSD for Steam and Epic stores
Display(s) 3 x 23" LG IPS panels (can't remember model!!)
Case 10mm thick MDF on plastic risers.. It's kinda a case??
Audio Device(s) Onboard
Power Supply EVGA P2 1200w
Mouse Corsair thingy
Keyboard Logitec K120......
Software Windows 10
Benchmark Scores I've actually never benched it!! Too busy with WCG and FAH and not gaming!! :( :(
Thank you for an excellent read :)

I do wonder if the CPU blocks for TR aren't quite big enough and judging by the pictures they could make them a little bigger to cover the CPUs better.. ? I'm definitely no expert but I do wonder if it will help :)
 

1usmus

AMD Memory Guru
Joined
Jul 28, 2018
Messages
61 (0.06/day)
Location
Ukraine
Thank you for an excellent read :)

I do wonder if the CPU blocks for TR aren't quite big enough and judging by the pictures they could make them a little bigger to cover the CPUs better.. ? I'm definitely no expert but I do wonder if it will help :)


You are absolutely right, soon the water blocks will be updated, as these water blocks are not perfect. This also applies to the camera area and depth of the micro-channels.
It may also be necessary to increase water resistance inside the unit for better turbulence.
 

phill

Moderator
Staff member
Joined
Jun 8, 2011
Messages
9,999 (2.75/day)
Location
Somerset, UK
System Name Not so complete or overkill
Processor Ryzen 3900X
Motherboard Asus Crosshair 8 Hero (Non Wi-Fi)
Cooling Air :(
Memory Patriot 3200MHz 64GB
Video Card(s) 2 x EVGA GTX 1080 Ti SC Black Edition
Storage 1 250GB Crucial SSD for OS/Programs, 2 500GB Crucial SSD for Steam and Epic stores
Display(s) 3 x 23" LG IPS panels (can't remember model!!)
Case 10mm thick MDF on plastic risers.. It's kinda a case??
Audio Device(s) Onboard
Power Supply EVGA P2 1200w
Mouse Corsair thingy
Keyboard Logitec K120......
Software Windows 10
Benchmark Scores I've actually never benched it!! Too busy with WCG and FAH and not gaming!! :( :(
You are absolutely right, soon the water blocks will be updated, as these water blocks are not perfect. This also applies to the camera area and depth of the micro-channels.
It may also be necessary to increase water resistance inside the unit for better turbulence.
Wow, didn't think I would be right :laugh: But....

I was wondering this as I have a few air coolers I bought to test and the plate that sits on top of the IHS was barely covering the CPU at all. The cooler was a Arctic Freezer 34 Duo, granted it was only a budget cooler but I thought it would have been made bigger for better contact, it would have been an even better cooler in my eyes.

That said, it was working brilliantly well with my 1700X and 2700, my 3900X not so much as temps were hitting 85C after a short CB R15 run (it definitely was short as 24 threads of Ryzen making it finish within 10 seconds of me clicking the button!! Bit different to my E-350 CPU but I digress....)
With all that said, even with a Noctua 14D cooler on it, the temps only dropped 5C under load but even then I still thought it was a little too high for running WCG which I plan to do, but that's another story :)

For this reason I did feel that I would prefer to have a Threadripper system over the 3950X CPU simply due to the temps side of things, plus with those extra cores and such, it would definitely be an even better system. Granted it's a very big price jump but... If you're running things 100% or close too a lot of it's life, I'd have thought that lower temps etc, would have only been a bonus for the hardware to last longer..

But it's just my thoughts and opinions, so what the heck do I know :D
 
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
Thanks for this excellent writeup.

One question though: Why is the overclocked setting (iOC in your case) slower than the stock setting in R20 Single? Especially why is stock + 3733-CL16 slower that stock + 3733-CL18?



Shouldn't it be the other way around?

I get that cheap overclocking with fixed multipliers contradict single-core boost on stock settings, but I guess that is not the explanation with your iOC settings as those should get higher scores across all workloads?!
 
Joined
Feb 26, 2016
Messages
416 (0.22/day)
Location
Texas
System Name O-Clock
Processor Intel Core i7-3770K @ 4.5 GHz delidded w/ Conductonaut
Motherboard ASUS Z77 Sabertooth
Cooling Enermax Liqmax II 240 push/pull w/ ICD thermal compound
Memory Mushkin Blackline 16GB (4x4GB) @2133 MHz CL10
Video Card(s) Asus GTX 1650 Super TUF OC @ +60 MHz/1750 MHz (core/memory)
Storage PNY CS1311 128GB 2.5" SSD, Samsung 860 QVO 1TB 2.5" SSD, Samsung 860 PRO 256GB 2.5" SSD
Display(s) Asus VG279QM, Acer G247HYL
Case Corsair SPEC-02 (modded)
Audio Device(s) HyperX Cloud II
Power Supply EVGA 1200 P2
Mouse Logitech G403
Keyboard Logitech G910 Stickerbombed
Software Windows 10 Pro 64 bit
Thanks for this excellent writeup.

One question though: Why is the overclocked setting (iOC in your case) slower than the stock setting in R20 Single? Especially why is stock + 3733-CL16 slower that stock + 3733-CL18?



Shouldn't it be the other way around?

I get that cheap overclocking with fixed multipliers contradict single-core boost on stock settings, but I guess that is not the explanation with your iOC settings as those should get higher scores across all workloads?!
When you manually overclock Zen2, as in, you put a fixed multiplier, the boosting mechanism is essentially disabled. The way boosting works for Ryzen (also applies to Intel but we will stick to Ryzen) is, when there is light loads or few threads loaded, the processor operates at a higher frequency with higher voltage, but the temps are no issue since it is usually sub 25% load when loading 1 or 2 threads. Therefore, the processor is actually stable under light loads for slightly higher frequencies. You can run offsets for voltages to allow for slightly higher headroom for boosting, but when you manually apply a fixed multiplier, then that is the maximum the processor will run at, and it must be stable for when ALL CORES/THREADS are loaded, not just a few. That is most likely why the overclocked results are lower than stock. Not to mention, Ryzen actually is extremely intelligent with the way they handle boosting, as it basically finds the most optimal frequency for normal users, however when running max load, it would be nice to also have it run at the best voltage/frequency (which is usually around 1.325V max and a variable frequency dependent on the silicon quality).

Hold up a minute, why is the 2990WX higher than the 3960X..?
 
Joined
Aug 8, 2019
Messages
428 (0.66/day)
System Name R2V2 *In Progress
Processor Ryzen 7 2700
Motherboard Asrock X570 Taichi
Cooling W2A... water to air
Memory G.Skill Trident Z3466 B-die
Video Card(s) Radeon VII repaired and resurrected
Storage Adata and Samsung NVME
Display(s) Samsung LCD
Case Some ThermalTake
Audio Device(s) Asus Strix RAID DLX upgraded op amps
Power Supply Seasonic Prime something or other
Software Windows 10 Pro x64
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.

A driver that boosts and outputs the same signal can't be called a doubler or balancer. It's an amplifier and even that simple action will delay the response slightly.

I was actually thinking about what it should be called to be honest but marketing friendly .

8 phases with doubled stages.
16 power stages running in an 8 phase configuration.

Asus VRM temp sensors are notoriously inaccurate from what I can find. They all seem to report 20-30'C low.

Honestly you are right an 8 phase single stage setup will run warmer than an 8x2 because of component surface area. It also makes the VRM twice as large so much easier to cool. There natural balancing at play so one stage isn't doing all of the work but it's not 50/50, the more stages the muddier things get.

If we look at current mobos in similar pricing ranges. IIRC...

X570 Prime by Asus, 3 actual phases running 12 stages for the CPU VRM.

X570 Aorus Gaming by Gigabyte, 6 actual phases into 12 phases running 12 stages, with the use of doublers. The doublers are intelligent and do active balancing between their connected phases.

X570 Aorus Pro by Gigabyte, 6 actual phases into 12 phases running 12 stages, using doublers. The doublers are dumb and the controller itself can control the balancing of the phases.

X570 Aorus Master 12 actual phases running 12 stages.

Guess which runs cooler on VRM testing? According to Hardware Unboxed, the Gigabytes. Even the Asus Hero gets smacked by the Aorus Master.
 
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
When you manually overclock Zen2, as in, you put a fixed multiplier, the boosting mechanism is essentially disabled. The way boosting works for Ryzen (also applies to Intel but we will stick to Ryzen) is, when there is light loads or few threads loaded, the processor operates at a higher frequency with higher voltage, but the temps are no issue since it is usually sub 25% load when loading 1 or 2 threads. Therefore, the processor is actually stable under light loads for slightly higher frequencies. You can run offsets for voltages to allow for slightly higher headroom for boosting, but when you manually apply a fixed multiplier, then that is the maximum the processor will run at, and it must be stable for when ALL CORES/THREADS are loaded, not just a few. That is most likely why the overclocked results are lower than stock. Not to mention, Ryzen actually is extremely intelligent with the way they handle boosting, as it basically finds the most optimal frequency for normal users, however when running max load, it would be nice to also have it run at the best voltage/frequency (which is usually around 1.325V max and a variable frequency dependent on the silicon quality).

Hold up a minute, why is the 2990WX higher than the 3960X..?

Yeah OK, I get that. However, the question was more like if it is possible to retain the individual-core-boosting feature while tweaking multipliers and number of cores to boost simulateously while keeping everything under a certain power and/or temperature limit? These are features you can tweak on Intel platforms which makes it a lot of fun to overclock. On Intel HEDT (FCLGA-2066 or even FCLGA-3647) you can even set a max. multiplier on each individual core (provided your mainboard manufacturer has that feature covered in the BIOS/UEFI of course).
I hoped this may be also available on TRX40. I mean on stock AMD shows that this is possible with their boosting scheme, I guess it isn't too far fetched to expect this to be tweakable, isn't it?
 
Joined
Feb 26, 2016
Messages
416 (0.22/day)
Location
Texas
System Name O-Clock
Processor Intel Core i7-3770K @ 4.5 GHz delidded w/ Conductonaut
Motherboard ASUS Z77 Sabertooth
Cooling Enermax Liqmax II 240 push/pull w/ ICD thermal compound
Memory Mushkin Blackline 16GB (4x4GB) @2133 MHz CL10
Video Card(s) Asus GTX 1650 Super TUF OC @ +60 MHz/1750 MHz (core/memory)
Storage PNY CS1311 128GB 2.5" SSD, Samsung 860 QVO 1TB 2.5" SSD, Samsung 860 PRO 256GB 2.5" SSD
Display(s) Asus VG279QM, Acer G247HYL
Case Corsair SPEC-02 (modded)
Audio Device(s) HyperX Cloud II
Power Supply EVGA 1200 P2
Mouse Logitech G403
Keyboard Logitech G910 Stickerbombed
Software Windows 10 Pro 64 bit
Yeah OK, I get that. However, the question was more like if it is possible to retain the individual-core-boosting feature while tweaking multipliers and number of cores to boost simulateously while keeping everything under a certain power and/or temperature limit? These are features you can tweak on Intel platforms which makes it a lot of fun to overclock. On Intel HEDT (FCLGA-2066 or even FCLGA-3647) you can even set a max. multiplier on each individual core (provided your mainboard manufacturer has that feature covered in the BIOS/UEFI of course).
I hoped this may be also available on TRX40. I mean on stock AMD shows that this is possible with their boosting scheme, I guess it isn't too far fetched to expect this to be tweakable, isn't it?
In a way yes. If I am not mistaken, there is a maximum boost override with PBO, and you can set that up to 200 MHz. However, I wonder if you could set a base frequency, and a max boost speed... Please forgive me for I personally do not own any Ryzen systems, but I have worked with Ryzen before, and I think the only real thing you could do (with Ryzen) is undervolt the processor, which has been shown to improve both single and multithreaded clocks, which improves those performances. I know what you are asking, and personally I wish Ryzen could implement a manual boost algorithm, similar to Intel's turbo boost ratio algorithm, except w/o the bogus power limits and stuff like that. THAT would be an overclocker's dream. Speaking of this, why can't ThrottleStop work with Ryzen? Is it possible to make a version for Ryzen, specifically for this, so that we could have the single core turbo, and a manual minimum clock speed set by user (for multithreaded workloads)? I have this on my laptop with multiple profiles, which results in exceptional battery life when not plugged in, and killer workstation performance when plugged in.

All in all, the only solution (that I can think of) that can actually increase the clocks on the boosting algorithm is by putting setting core voltage to offset mode and put an offset. Oh also, don't use a Shittabyte motherboard. I know why they have problems with any sort of tuning, it's because their "genius" engineers thought it was a good idea to provide VRMs that do not support voltage control, that is why voltage is usually funky on Gigabyte boards. If you go Asus or MSI or ASRock, you are bound to have a board that will actually be capable of *changing* the voltage from stock.
 
Joined
Jun 2, 2017
Messages
3,762 (2.60/day)
System Name Best AMD Computer
Processor AMD 5600X
Motherboard MSI X570 Unify
Cooling Silverstone IceGem 280
Memory ADATA D60 3600MHZ 32 GB
Video Card(s) MSI 6800XT Gaming Trio X
Storage Seagate 520 1Tb, Adata SX8200 2TBx2, Kingston 2 TBx2, Micron 8 TB
Display(s) GIGABYTE 32QC
Case PHANTEKS ENTHOO PRO 2
Audio Device(s) Corsair Void Pro, Logitch Z523 5.1
Power Supply Corsair HX1200!
Mouse Logitech g7 gaming mouse
Keyboard Logitech G510
Software Windows 10 Pro 64 Steam. GOG, Uplay, Origin
Benchmark Scores Firestrike: 24955 Time Spy: 13500
A driver that boosts and outputs the same signal can't be called a doubler or balancer. It's an amplifier and even that simple action will delay the response slightly.

I was actually thinking about what it should be called to be honest but marketing friendly .

8 phases with doubled stages.
16 power stages running in an 8 phase configuration.

Asus VRM temp sensors are notoriously inaccurate from what I can find. They all seem to report 20-30'C low.

Honestly you are right an 8 phase single stage setup will run warmer than an 8x2 because of component surface area. It also makes the VRM twice as large so much easier to cool. There natural balancing at play so one stage isn't doing all of the work but it's not 50/50, the more stages the muddier things get.

If we look at current mobos in similar pricing ranges. IIRC...

X570 Prime by Asus, 3 actual phases running 12 stages for the CPU VRM.

X570 Aorus Gaming by Gigabyte, 6 actual phases into 12 phases running 12 stages, with the use of doublers. The doublers are intelligent and do active balancing between their connected phases.

X570 Aorus Pro by Gigabyte, 6 actual phases into 12 phases running 12 stages, using doublers. The doublers are dumb and the controller itself can control the balancing of the phases.

X570 Aorus Master 12 actual phases running 12 stages.

Guess which runs cooler on VRM testing? According to Hardware Unboxed, the Gigabytes. Even the Asus Hero gets smacked by the Aorus Master.

I personally will never buy a GIgabyte board ever again. Asus is not far behind either. As rock are for me the best MBs for Ryzen based systems.
 
Joined
Jul 10, 2019
Messages
12 (0.02/day)
System Name Waterbucket
Processor AMD Ryzen 9 3950X (Thermal Grizzly Kryonaut)
Motherboard Gigabyte X570 Aorus Xtreme
Cooling Aquaero 6 PRO; D5+PMP-450; EK-Velocity; EK-FC Titan X; Alphacool Rad. 480 (60 & 80) + NF-F12 PWM x16
Memory G.Skill TridentZ Neo 3600 CL16 (2x16GB)
Video Card(s) EVGA GTX 980 Ti SC
Storage Corsair MP510 960GB x2; Samsung 850 PRO 512GB; Crucial M550 256GB; WD VelociRaptor 1TB; WD30EFRX x2
Display(s) Asus PG2749Q + Dell UltraSharp U3011
Case Thermaltake Core W200
Audio Device(s) MOB + Logitech Z5500
Power Supply SeaSonic Platinum-1200 XP3
Mouse Logitech G602
Keyboard SteelSeries Apex M800
Software Win 10
A very nice article. Factfully and quite well written.

Is there a chance for the OC Guide for regular Ryzen 3000? (3900X, 3950X)
 

W1zzard

Administrator
Staff member
Joined
May 14, 2004
Messages
22,108 (3.56/day)
Processor Core i7-8700K
Memory 32 GB
Video Card(s) RTX 3080
Display(s) 30" 2560x1600 + 19" 1280x1024
Software Windows 10 64-bit
Is there a chance for the OC Guide for regular Ryzen 3000? (3900X, 3950X)
 
Joined
Aug 8, 2019
Messages
428 (0.66/day)
System Name R2V2 *In Progress
Processor Ryzen 7 2700
Motherboard Asrock X570 Taichi
Cooling W2A... water to air
Memory G.Skill Trident Z3466 B-die
Video Card(s) Radeon VII repaired and resurrected
Storage Adata and Samsung NVME
Display(s) Samsung LCD
Case Some ThermalTake
Audio Device(s) Asus Strix RAID DLX upgraded op amps
Power Supply Seasonic Prime something or other
Software Windows 10 Pro x64
I personally will never buy a GIgabyte board ever again. Asus is not far behind either. As rock are for me the best MBs for Ryzen based systems.

I've had some issues with Gigabyte, but I just used their boards as a comparison. Not nearly the same level of never again with Asus though. LoL

I've had some luck with Asrock overall.

The Asrock X570 TaiChi/Phantom Gaming X/Creation use the same VRM as far as I know as the Aorus Gaming.

I think the important thing is to remember every mobo maker makes bad boards and good boards. Asus is on the down swing and Gigabyte on the up. Personally I think Gigabyte is building the best boards of the high end for the X570/TRX40 currently.
 

Kicksta

New Member
Joined
Feb 12, 2020
Messages
2 (0.00/day)
Excellent article, one of the reasons I consider this site one of the top in its' field.
 
Joined
Jul 10, 2019
Messages
12 (0.02/day)
System Name Waterbucket
Processor AMD Ryzen 9 3950X (Thermal Grizzly Kryonaut)
Motherboard Gigabyte X570 Aorus Xtreme
Cooling Aquaero 6 PRO; D5+PMP-450; EK-Velocity; EK-FC Titan X; Alphacool Rad. 480 (60 & 80) + NF-F12 PWM x16
Memory G.Skill TridentZ Neo 3600 CL16 (2x16GB)
Video Card(s) EVGA GTX 980 Ti SC
Storage Corsair MP510 960GB x2; Samsung 850 PRO 512GB; Crucial M550 256GB; WD VelociRaptor 1TB; WD30EFRX x2
Display(s) Asus PG2749Q + Dell UltraSharp U3011
Case Thermaltake Core W200
Audio Device(s) MOB + Logitech Z5500
Power Supply SeaSonic Platinum-1200 XP3
Mouse Logitech G602
Keyboard SteelSeries Apex M800
Software Win 10
There is nothing about OC processor :\

I can OC memory. It's worse with the CPU.
 

Asni

New Member
Joined
Feb 14, 2020
Messages
5 (0.01/day)
Hi @1usmus ,first at all i would like to thank you for your guide.
I read it a couple of time but i have some doubts about two steps:
- if your OC in CCX based how could you find the voltage required for every single core?
- given the target, 1.35v 4.4ghz, why did you reduce the clock of some CCX? If i read your core map correctly every single core can do 4.4ghz within the limit. Am i wrong?

Thanks in advance.
 
Joined
Sep 25, 2019
Messages
61 (0.10/day)
Processor AMD Threadripper 2950X
Motherboard ASUS ROG Zenith Extreme
Cooling Custom hard-line water HeatKiller Pro CPU+VRM, Dual Radiator, EK Res'/fittings
Memory 64GB Corsair Dominator RGB
Video Card(s) Liquid Devil 6900XT
Storage Samsung 512GB Pro M.2. Intel 1.2TB PCI NVMe, 3 X 512GB SSD
Display(s) ASUS ROG 34" Curved UW 3440x1440 100Hz
Case Thermaltake View 71T (10 Fans)
Power Supply Corsair HX1000i
Mouse Logitech G502
Keyboard ASUS ROG
Software Win 10 PRO x64
Benchmark Scores Cinebench R20: 7984
A great read - and bookmarking this review for a later date when I get my 3960x (or equiv').

It felt as though there were a couple of good, high-level, take aways from this.

1) The 3960x, when 'tuned', can receive a serious bump in gaming performance.
2) It was already competitive, but with a serious bump in gaming performance, it will knock more 9900K(KS) number 1's into second spot (unverified).
3) It will differentiate those that think they are overclockers, and those who really are.

Yes it's expensive for gaming, but there'll be a lot more people that buy one if they think it has the potential to become the new gaming king. After all, these people are 'serious' overclockers - right?
 
Last edited:
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
@1usmus: I tried to replicate your results myself with my 3960X (now I have one at last). May I ask what version of Prime95 you were using?
I tried this with the latest v29.8, build 6 and find the "Small FFT" (36K..248K) absolutely horrendously demanding in comparison to "Smallest FFT" (4K..21K) or "Large FFT" (426K..8192K).

For instance with my "intelligentOC" the CPU alone is pulling ~380W in Large FFT, but around 550W with Small FFT. The system is pulling close to 800W at the wall at this point and that is without GPU usage. This difference is surprisingly huge and no matter what other benchmark/stress-test I ran, I can not even come close to what Prime95 in Small FFT mode is pulling. So I wonder if it does make sense to tune the OC for that knowing that I will not reach such levels in any real world application?!

Allthough I have a full custom water cooling with a total of four 480mm 65mm thick radiators in push-pull and an EK monoblock with liquid metal on the CPU, this is simply too much for ambient cooling. The CCDs are reaching 100°C with that while on the less demanding Large FFT they sit in the high-60s.
 
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
Hi @1usmus ,first at all i would like to thank you for your guide.
I read it a couple of time but i have some doubts about two steps:
- if your OC in CCX based how could you find the voltage required for every single core?
- given the target, 1.35v 4.4ghz, why did you reduce the clock of some CCX? If i read your core map correctly every single core can do 4.4ghz within the limit. Am i wrong?

Thanks in advance.

This is probably because all-core load is much harder for the CPU. For instance you can get a voltage/frequency setting stable on one core or even on all threads in one CCX, but when you apply that to all threads, it isn't stable anymore. I've found that you have to back the clocks off by about 50-100MHz when testing all-core compared to single-CCX on the same voltage. Further testing all-core produces much more heat and if your cooling solution can't keep up, then you have to back down the voltages and start over.
 

froboy88

New Member
Joined
Aug 7, 2020
Messages
2 (0.01/day)
@1usmus, Thank you for this detailed write up on the 3960x. Most, as you said, dismissed overclocking this CPU out of hand and as such, there isn't much info out there about it.

You mentioned in your post the power draw at your peak OC was 337W. Did you take note of the power draw at lower OC? Say 4-4.2ghz?

I'm trying to plan a custom loop for the 3960x and dual GPU's and am finding it difficult to work out exactly how much load the CPU could have for a more modest OC. You're article cleared up more then any other thing I have read but I still have some questions.

How did the CE420's manage the coolant temp? You show how much them can affect the max OC so I had to assume that they manage to keep the core temp around 50C.

Did you do any tests Loading the CPU and GPU? Were the CE420's still enough in that workload?

I'm trying to work out if dual PE360's will be enough for a more modest OC. In your experience what do you think they could handle? With Dual 2080ti's and a work load that hits everything hard like video rendering would more radiator be needed?

Thanks
 
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
I guess at some point more radiator space is not helping to get the temperature of this CPU low enough.

I have a 3960X and have it overclocked to mostly 4.4 with the exception of 3 CCXes, which I only get stable at 4.325, 4.350 and 4.375 respectively.
I have a custom loop with four thick 480 rads (Aquacomputer AMS) with push-pull fans.

At this point rad space is not the issue, as the heat transfer at the CPU block is simply not good enough to efficiently get the heat away from the CPU and into the water. I'm using liquid metal and an EK block, but still am reaching 85°C core temp using Prime95 while pulling close to 500W CPU alone. Water temp stays below 35°C on the cool side, even with ambient temp around 25°C. So there is a delta of at least 50K on the CPU there.

Maybe that gives you an idea about what to expect on the upper level. At stock this chip pulls about 280W sustained under load, so depending on your overclock you will land somwhere in-between.
 

froboy88

New Member
Joined
Aug 7, 2020
Messages
2 (0.01/day)
@Breit, Thanks for your input. I'm surprised, that with that much radiator, your coolant is still 10C over ambient but I guess, in general, radiators can't do much without a delta between ambient.

Back in the day, I decided my Ivybridge 3770k, because like the threadripper it didn't matter how much cooling you threw at it, temps spiked in the CPU as there was too much heat trapped in the IHS. Have you seen the video where der8auer delids his 3970x? The chip is so much more then the 3770k I don't think I would try it even if it helped.

Have you monitored temps outside of Prime? While 85C is high given your coolant temp, its not bad for Prime in general, I just wouldn't want to run it 24/7. Do you render or compile with the chip?

If I read @1usmus charts correct, the 3960x will boost to 4.2GHZ, if the chip stays under 50C. Is this an all-core boost clock? Have you let the 3960x run stock? What kind of boost do you see with your setup?

Regarding radiators: I'm wondering how well 1 or 2 of your 480's would do as I'm interested in running either 2x360 or 1x360 + 1x420. I'd expect that, in push-pull, they can do over 500w/10C over ambient (correct me if I'm wrong). Does that mean that with your coolant temp your setup has about 2000W of cooling power?

I'm trying to find the sweet spot for Rads, and it seems like with 2x GPU's and a 3960x the total TDP might be over 1200W (500W+380Wx2). However, I may take a more optimistic approach, and either use a lower OC or account for few workloads using both CPU and GPU to 100%. "Optimistic" loop TDP might be 330W x 3 =990W. This might still leave room for either the GPU, or CPU to not be thermaly constrained, just not both at the same time. As I said I can't think a non-synthetic I would use that would hit both at 100%

What are your thoughts on the minimum radiator size for OC that setup? It looks like an EK CE420 and XE360 both Push/pull could pull that off but not sure if that even fits. I'm trying to stick to a Define S2 or Define 7 size case. Another option would be CE420 + PE360 + 140/180 all push.
 
Joined
Jun 4, 2004
Messages
455 (0.07/day)
System Name Blackbird
Processor AMD Threadripper 3960X 24-core
Motherboard Gigabyte TRX40 Aorus Master
Cooling Full custom-loop water cooling, mostly Aqua Computer and EKWB stuff!
Memory 4x 16GB G.Skill Trident-Z RGB @3733-CL14
Video Card(s) Nvidia RTX 3090 FE
Storage Samsung 950PRO 512GB, Crusial P5 2TB, Samsung 850PRO 1TB
Display(s) LG 38GN950-B 38" IPS TFT, Dell U3011 30" IPS TFT
Case CaseLabs TH10A
Audio Device(s) Edifier S1000DB
Power Supply ASUS ROG Thor 1200W (SeaSonic)
Mouse Logitech MX Master
Keyboard SteelSeries Apex M800
Software MS Windows 10 Pro for Workstation
Benchmark Scores A lot.
I'd say with 2 rads (preferably 420 or 480) you should be fine, even with 2 GPUs. The CPU and maybe the VRM alone will be fine with only one rad. I also have 2 GPUs which pull ~300W each under load. I only have 4 rads because I mostly wanted a silent PC (idle and light load, it is whisper quiet) and they fited without problem in my case.

On stock my chip boosts to 4500-4550 and VERY occasionally 4600. But sustained non-Prime95 load it stays around 4000 and with prime it drops to 3850-3900.
When I render stuff or encode videos (100% load requires multiple streams to be rendered simultaneously) the CPU stays around 60°C with my overclock, only prime95 is so demanding. I could've overclocked the chip to even 4500 and two CCXs to 4600 if prime95 wouldn't be so hard.
 
Top