PrimeGrid PPS LLR (CPU temp) challenge (for Ryzen 3000-series owners)! (Testing heat density of TSMC 7nm)

[john3850]

Were does Amd get the R5 3600 from could be that they were 3700 with too high of a TDP. I always believed the slower Intel and Amd chips were just defected higher binned chips and when they run out they use higher binned chip which are good overclocking chips.

 

[DrMrLordX]

I went and tried PPS LLR myself. It's a bit toasty folks! Tried my Prime95 Small FFTs settings (1.25v, 4200-4250 MHz) and the thing shot up in temps like mad, close to 80C. This definitely is the hottest workload I've ever tried.

BUT

Right now I've got 4100 4150 MHz with 1.175v running just fine. 69C, and it's stable. Actual voltage is closer to uhhh hold on. 1.212v thanks to Turbo LLC. Gotta love it. (edit: actually fluctuating between 1.212 and 1.200v but whatever).

Anyway

@VirtualLarry

Just because you are seeing the chip feed that much voltage in by default doesn't mean you want to select that voltage in a static overclock. Don't pick voltages that high, especially not in an AVX2 workload. Recipe for disaster.

 

[therealmongo]

So im a little late,

here are my temps, custom water

 

[Zstream]

Yeah voltages tend to run high. My stock is around 1.3v. It definitely doesn’t need to be that way. I’ll tune it down later this week, but yes, don’t keep that stock voltage. I haven’t tested that software, but cinebench never goes above 70c and my ambient temp is 75F-76F

 

[therealmongo]

Yeah voltages tend to run high. My stock is around 1.3v. It definitely doesn’t need to be that way. I’ll tune it down later this week, but yes, don’t keep that stock voltage. I haven’t tested that software, but cinebench never goes above 70c and my ambient temp is 75F-76F

I have it on PBO 10x and boost of 75mhz, voltages on auto.

Lets see how long my processor lasts!

 

[DrMrLordX]

Lets see how long my processor lasts!

FOR SCIENCE

 

[therealmongo]

FOR SCIENCE

 

[StefanR5R]

I went and tried PPS LLR myself. It's a bit toasty folks! Tried my Prime95 Small FFTs settings (1.25v, 4200-4250 MHz) and the thing shot up in temps like mad, close to 80C. This definitely is the hottest workload I've ever tried.

Regarding @VirtualLarry's original request:

The LLR program which PrimeGrid uses, and Prime95, both use the same FFT library. Prime95's FFT size can be configured directly, while LLR chooses the FFT size automatically based on processor architecture (leads to same choice on all more or less current architectures as far as I have seen) and, more importantly, based on prime number search space. PrimeGrid's PPS subproject currently uses a relatively small FFT size of 192K.

Hence, in order to replicate @VirtualLarry's workload, Prime95 can be used too, if FFT size is chosen similarly, all cores are used, and the program is running for long enough.

Considering I can do 4.2GHz AVX2 stable on my worst chiplet and 4.35GHz AVX2 stable on my better chiplet at 1.25V those voltages are absurd for 4GHz.

But hey, don't take my word for it, Silicon Lottery's absolute worst binned 3900X is guaranteed stable under their conservative binning at 4GHz all-core at 1.20V.
https://siliconlottery.com/collections/matisse/products/3900x40g?variant=29472043499606

1.365V (and it looks like 1.40V is being applied from Larry's screenshots) is absurdly high for 4GHz manual OC on Ryzen 3000.

Right now I've got 4100 4150 MHz with 1.175v running just fine. 69C, and it's stable. Actual voltage is closer to uhhh hold on. 1.212v thanks to Turbo LLC. Gotta love it. (edit: actually fluctuating between 1.212 and 1.200v but whatever).

If voltage is marginal relative to what is really needed for a given core clock, so-called clock stretching gets activated. In effect, this clocks the cores down but core clock monitoring programs don't show it. It can be seen directly by monitoring a specific processor performance counter, or indirectly by measuring application performance.

Source: "Strictly technical: Matisse (Not really)" at OCN, e.g. post #229

Edit, therefore I suspect that many people who are seeing high clocks with non-default settings, especially with capped voltage, have clock stretching active, and either did not measure their application performance in comparison with stock settings, or did so and found that performance does not scale at all with (merely apparent!) core clocks.

 

[DrMrLordX]

If voltage is marginal relative to what is really needed for a given core clock, so-called clock stretching gets activated. In effect, this clocks the cores down but core clock monitoring programs don't show it. It can be seen directly by monitoring a specific processor performance counter, or indirectly by measuring application performance.

Source: "Strictly technical: Matisse (Not really)" at OCN, e.g. post #229

Edit, therefore I suspect that many people who are seeing high clocks with non-default settings, especially with capped voltage, have clock stretching active, and either did not measure their application performance in comparison with stock settings, or did so and found that performance does not scale at all with (merely apparent!) core clocks.

I've seen clock stretching. You actually can't get clock stretching to kick in with a static OC unless you pick a speed below the base. For speeds above base (3.8 GHz for me), supplying too little voltage just causes the machine to crash. I've tested it extensively with applications that give handy performance-based readouts like CBR20. With PB2 + XFR you can get clock stretching at any clockspeed, usually with negative voltage offsets in the UEFI, though that's not the only way you'll see it.