Question 3950x Prime95 Clock speed

[relmik]

Hi All,

I have spent the last few days searching these forums looking for answers and have found a bunch of great info. But i'm still a little perplexed about the problem that I am having and hope you all can shed some light on it. This is my first build with a Ryzen processor so im pretty noobish here.....

First things first, my computer build:
CPU: Ryzen 3950x
Cooler: Noctua NH-D15 (using both fans, ARCTIC MX-4 Thermal Paste)
Motherboard: Gigabyte Ultra X570
Memory: G.Skill Trident Z Neo 32gb, F4-3600C16D-32GTZNC
Case: Fractal Design Define R6 (7 fans total pushing air from front to back and bottom to top)
PSU: Corsair RM850x

Current Tweaking:
-I have the latest BIOS and chipset drivers installed.
-XMP is on with memory at 3600
-PBO is off and LLC is High
-Vcore set to 1.325

Im not trying to set world records with this build/setup. Just wanted to futureproof myself for gaming as I only upgrade every 6-8 years or so. However, I always like to run Prime95 just to check thermals and make sure everything is ok. So I ran 3 tests on Prime95 and have attached screenshots of each. 1) Blend test, AVX off running 16 threads. 2) Small FFT, AVX off running 16 threads 3) Small FFT, AVX off running 32 threads. All screenshots taken within the first 30 seconds of initiating the Prime95 runs.

With the current BIOS setup, Everything looks good in the Blend test (at least to me it does). However, in both the Small FFT runs, im throttled all the way down to 2.9 Ghz and 2.4 Ghz. At these speeds my temps still look decent. If I let the Small FFT's run longer I will eventually get to 80-82 deg C.

So what gives on the clock speed limitation right out of the gate? Temps seem fine so why such a dramatic reduction in clock speed? Is this a voltage drop issue or possibly power/current limitation in a setting somewhere in the BIOS? I appreciate any help yall can offer up.

 

[.vodka]

Current Tweaking:
-I have the latest BIOS and chipset drivers installed.
-XMP is on with memory at 3600
-PBO is off and LLC is High
-Vcore set to 1.325

Is 1.325v your FIT voltage? If not, you're gonna degrade that poor 3950x. You should be aiming for ~1.2 - 1.25v if you want a fixed manual all core OC on the dual CCD parts (3900x, 3950x). Having said that, you're not going to do better than the boost algorithm on Zen2/3000 series CPUs. Disable PBO, set LLC to auto, set vCore to auto. Let the thing manage itself.

By itself it'll pull up to 142w total package power unless you activate PBO or raise the PPT/TDC/EDC limits which are set to 142W, 95A, 140A.

Use hwinfo. You'll be able to see much better what's going on. Hwmonitor is best suited for old platforms.

An effective tweak on 3000 series CPUs is a small general negative offset undervolt (-0.05 to 0.1v), this will free some power in the PPT budget representing a sacrifice of 1-2% single core boost for a larger multithreaded boost on >4 threads, and a ~10°C temp decrease if -0.1v is stable. Undervolt and run some benchmarks, once you start losing performance that's the sign you're going too far. If you keep going it'll just crash.


Anyway, P95 is a power virus type of workload, any modern CPU will throttle hard on it. Yours throttles harder because 1.32v is an insanity for such a workload. See, the boost algorithm also scales voltage back on such workloads as to keep per core power down and clock speeds up. P95 small FFTs on 3000 series CPUs will usually result in a voltage in the region of 1-1.1v, not 1.32v, and clock speeds near the specified base clock, not lower.

See how little it throttles once you've undone your tweaks and let it manage itself.

 

[relmik]

Is 1.325v your FIT voltage? If not, you're gonna degrade that poor 3950x. You should be aiming for ~1.2 - 1.25v if you want a fixed manual all core OC on the dual CCD parts (3900x, 3950x). Having said that, you're not going to do better than the boost algorithm on Zen2/3000 series CPUs. Disable PBO, set LLC to auto, set vCore to auto. Let the thing manage itself.

By itself it'll pull up to 142w total package power unless you activate PBO or raise the PPT/TDC/EDC limits which are set to 142W, 95A, 140A.

Use hwinfo. You'll be able to see much better what's going on. Hwmonitor is best suited for old platforms.

An effective tweak on 3000 series CPUs is a small general negative offset undervolt (-0.05 to 0.1v), this will free some power in the PPT budget representing a sacrifice of 1-2% single core boost for a larger multithreaded boost on >4 threads, and a ~10°C temp decrease if -0.1v is stable. Undervolt and run some benchmarks, once you start losing performance that's the sign you're going too far. If you keep going it'll just crash.


Anyway, P95 is a power virus type of workload, any modern CPU will throttle hard on it. Yours throttles harder because 1.32v is an insanity for such a workload. See, the boost algorithm also scales voltage back on such workloads as to keep per core power down and clock speeds up. P95 small FFTs on 3000 series CPUs will usually result in a voltage in the region of 1-1.1v, not 1.32v, and clock speeds near the specified base clock, not lower.

See how little it throttles once you've undone your tweaks and let it manage itself.

Thank you for all the great info. I had read where quite a few people were using 1.3 to 1.35 for the vcore, I didnt know that was unsafe. I thought that was just a cap, limiting the CPU from exceeding that voltage and would only be possibly bad if sustained under load for long periods of time. Im only asking because im curious and want to better understand whats going on.

Is VDDP the undervolt setting you are talking about? Still getting use to this motherboard BIOS.

Oh and one other thing, I didnt think I had anything overlocked other than just my RAM. PBO is off and I havent messed with anything else other than just the LLC.

 

[relmik]

I made the changes that you recommended including an offset of -0.05v.

Here is what im seeing now:
-At idle, HWinfo is showing the voltage on the cpu to be at 1.35-1.4 volts. So it went up from my 1.325 v. But this is only at idle. So that surprised me. Vcore is Auto in BIOS.
-When I run Prime95 SmallFFT, the voltage now drops to 1.02. So sounds like this was to be expected as it is reducing the voltages to keep the clocks up and heat down.
-I am dropping two worker windows while running Prime95 now. I receive the error "FATAL ERROR: Resulting sum was 6.443957013765818e+35, expected: 6.443957013765829e+35"

So let me know if I understand this correctly... The higher voltage may not be an issue since that is only at idle. Once a load is placed on the CPU, the voltage is regulated down. Temps never got above 74C. Is this how it should be operating? I was originally thinking I would never see the voltage above 1.2-1.25 .

I don't know what to think that that Prime95 error. Do I need to increase the under voltage (positive direction) and try to find the sweet spot? Or just give up on Prime 95? lol

 

[.vodka]

I made the changes that you recommended including an offset of -0.05v.

Here is what im seeing now:
-At idle, HWinfo is showing the voltage on the cpu to be at 1.35-1.4 volts. So it went up from my 1.325 v. But this is only at idle. So that surprised me. Vcore is Auto in BIOS.
-When I run Prime95 SmallFFT, the voltage now drops to 1.02. So sounds like this was to be expected as it is reducing the voltages to keep the clocks up and heat down.
-I am dropping two worker windows while running Prime95 now. I receive the error "FATAL ERROR: Resulting sum was 6.443957013765818e+35, expected: 6.443957013765829e+35"

So let me know if I understand this correctly... The higher voltage may not be an issue since that is only at idle. Once a load is placed on the CPU, the voltage is regulated down. Temps never got above 74C. Is this how it should be operating? I was originally thinking I would never see the voltage above 1.2-1.25 .

I don't know what to think that that Prime95 error. Do I need to increase the under voltage (positive direction) and try to find the sweet spot? Or just give up on Prime 95? lol

Now that's more in line with what's expected.

These CPUs behave more like current GPUs (think of nV's GPU boost 3.0/4.0 based GPUs, AMD's Vega and Navi cards), in which clockspeed is highly variable and highest when the workload is light, pushing enough voltage to sustain that clockspeed while staying inside the operating parameters (PPT, TDC, EDC). Take for example the 3900X as it behaved a few BIOS releases ago:

PPT is total package power, TDC is maximum socket current, EDC is maximum peak current. The boost algorithm operates within these constraints and will never exceed them.

Voltage range for these CPUs is 0.2v idle (cores off) - 1.5v (yes, 1.5v). Have a look at the usual ranges for different use cases here. You are going to see 1.4-1.5v readings at idle because that's how it's supposed to work.

See how it's a decreasing curve? When the boost algorithm is pushing 1.4-1.5v it's through one core that's running at maximum speed (4.7GHz if needed in case of the 3950x) WHILE requiring low current to do so. Load more cores with heavier workloads that require more current, voltage goes down, frequency goes down following that curve. You're always getting maximum performance possible from the CPU, you just have to keep it cool as best as you can.

A negative voltage offset pulls the entire voltage curve down and it's the correct way to undervolt these CPUs. Sadly 0.05v is too much of an undervolt for your 3950X, you shouldn't be seeing P95 fail.

Now if you fail P95 without that undervolt, something's unstable in your system. Probably your memory running at 3600MHz needs a bit more tweaking than just loading the XMP profile and calling it a day.

Are you running the latest BIOS for your motherboard? You should be. Always do.

 

[relmik]

Thank you for the great write up! Now that it is all starting to come together, I didn't realize how far i was really pushing this CPU.

So I am running the latest BIOS. I still have the same Prime95 failure even turning undervolting off. So I adjusted my memory clock down to just under 3200. As long as I'm below that, everything runs good and I don't exceed 68 deg C during the thermal testing which is the best I have seen yet.

So, in order to get the memory up to 3600, what else can I adjust? The DRAM is set to auto. Should I change that to the rated voltage or am I trying to push the limits again here? Its rated to 1.35V.

 

[thigobr]

You should set Dram voltage to 1.35V (XMP should take care of that though).

If Prim95 is unstable with memory at 3600MHz but it's working with memory at 3200MHz it could be 2 things: memory needs more voltage/some timings tune or it could also be the memory controller needing more volts (Vsoc, VDDG, and VDDP voltages). If you can share a ZenTimings screenshot with memory set to 3600MHz we would be able to help you better.

ZenTimings

ZenTimings is a simple and lightweight app for monitoring memory timings on Ryzen platform.

zentimings.protonrom.com

 

[relmik]

Here is a screenshot while the computer is sitting idle.

 

[.vodka]

Here is a screenshot while the computer is sitting idle.

Manually set these voltages:

vSOC 1.1v
vDDG (both) 0.95v
CLDO_VDDP 0.9v

The rest looks fine as it is, for any further tweaking first we need to be stable at XMP settings. That should do it.

-----------------------

This is from your motherboard's manual.

Have you installed your two modules in slots A2 and B2? If not, please do.

 

[relmik]

Is vSOC the same as VCORE SOC?

I have the two memory sticks in the A2 and B2 slots.