PPD Database - HFM exports needed

[NoMoreQuarantine]

I would say there is nothing clear about the 2060 Super. If you look at the HFM database there is only a sample size of 1.

Ah, that changes things. So my prediction is more or less correct... until we have a larger sample set

[lafrad]

My all core-overclock is locked in. Cooling is great. power use of the chip is... high. I game and work on it, and Linux/Windows seem to be perfectly happy sharing FAH and whatever I'm doing on it too. without impact to the interactive task, and only (usually) a minor dip to the FAH PPD.

I'd just prefer that FAH itself isn't reducing the real compute value when the computer is not busy with other tasks.

[foldinghomealone]

I would say there is nothing clear about the 2060 Super. If you look at the HFM database there is only a sample size of 1.

Ah, that changes things. So my prediction is more or less correct... until we have a larger sample set

I would say that there is nothing clear at all, even with larger sample sizes.

No indication of hw and electricity costs and power consumption.

Even then, how to define and measure power consumption correctly?
- using TDP is useless
- power consumption varies over following factors: PT, PL, OC, UV and WU (different WUs utilize the GPU differently)
- measuring power consumption of a GPU is not easy - only possible with lab grade equipment and would be a tremendous effort for long term measurements
- using SW like GPU-Z give only indications

It would be easy to measure the system consumption including GPU, CPU, MB, HDD, SDD, cooling solution.
But then it's not easy to determine the power consumption of the GPU, either:
- one or more GPUs?
- CPU folding as well? How many threads?

Even not accurate, for most practical reasons, using SW like GPU-Z would be the way to go, I guess.
However, long term readings would be necessary to get accountable data

[foldinghomealone]

New data for:
R7 3700X
GTX 1660 Ti

RTX 2060 KO

Thanks to:
NoMoreQuarantine

[NoMoreQuarantine]

I would say that there is nothing clear at all, even with larger sample sizes.

No indication of hw and electricity costs and power consumption.

Even then, how to define and measure power consumption correctly?
- using TDP is useless
- power consumption varies over following factors: PT, PL, OC, UV and WU (different WUs utilize the GPU differently)
- measuring power consumption of a GPU is not easy - only possible with lab grade equipment and would be a tremendous effort for long term measurements
- using SW like GPU-Z give only indications

It would be easy to measure the system consumption including GPU, CPU, MB, HDD, SDD, cooling solution.
But then it's not easy to determine the power consumption of the GPU, either:
- one or more GPUs?
- CPU folding as well? How many threads?

Even not accurate, for most practical reasons, using SW like GPU-Z would be the way to go, I guess.
However, long term readings would be necessary to get accountable data

I agree, I was just joking around. You are completely correct about TDP being useless; the 2080 Ti has a rated TDP of 250W, but this 2080 Ti video card, as an example, has a power limit of 366W! If we were to have users submit the exact video card model with their data, we could cross reference it, but that is a big ask and a big headache. It's even more difficult with CPUs where the only power limit is the motherboard VRM (I am oversimplifying). I currently don't see a solution other than to automatically record system configuration and status with a complicated software solution (which I suspect is relatively accurate on most modern cards). That was a long way of saying I don't have any better ideas.

[lafrad]

Pretty sure the general idea for this was... to get a general idea. take the data from a spectrum of situations, and give an idea where you expect to be. Things as subtle as the brand of case fan will adjust the PPD of an air-cooled GFX card..... why worry about that point?

[Juggy]

@skydivingcatfan very interesting chart! Not the results I was expecting; I thought the 2070 Super would be in the lead the first couple years, but it looks likes the 2060 Super is the clear winner.

I would say there is nothing clear about the 2060 Super. If you look at the HFM database there is only a sample size of 1.

I also think that sample PPD is high in comparison to what I was seeing on a 2060 Super Gaming OC when I ran ~50 WU's through mine. Unfortunately I got rid of it before I installed HFM (I rebuilt my machine so lost all logs)

[Juggy]

New data for:
R7 3700X
GTX 1660 Ti

RTX 2060 KO

Thanks to:
NoMoreQuarantine

How many samples do you require for the dataset?

I submitted more for the i9 and 2080 Super. Total now of over 200 for 2080 Super and over 300 for i9

[NoMoreQuarantine]

Pretty sure the general idea for this was... to get a general idea. take the data from a spectrum of situations, and give an idea where you expect to be. Things as subtle as the brand of case fan will adjust the PPD of an air-cooled GFX card..... why worry about that point?

Yep, I am pretty happy with the current data. I feel it gives a good idea of how someone could expect a processor to perform, and supplies the raw data that can be used to perform analysis across the different projects and atom counts.

What it doesn't tell us is how energy efficient or cost effective the individual hardware is. Which are separate practical issues many of us are interested in.

[foldinghomealone]

How many samples do you require for the dataset?

That's a good question.
My first thought would be the more the better, however if there are too many samples of the same user / the exactly same CPU/GPU then this could be misleading.
Like one uses an UVed GPU and submits 1000 samples. Other samples from other users would have no influence on the average.

I would say if there are more than 100 samples of the same user it should be sufficient for now.
Maybe in a few weeks with new projects, new samples of the same user can be submitted.
What do you think?

[Juggy]

How many samples do you require for the dataset?

That's a good question.
My first thought would be the more the better, however if there are too many samples of the same user / the exactly same CPU/GPU then this could be misleading.
Like one uses an UVed GPU and submits 1000 samples. Other samples from other users would have no influence on the average.

I would say if there are more than 100 samples of the same user it should be sufficient for now.
Maybe in a few weeks with new projects, new samples of the same user can be submitted.
What do you think?

Yeah, it is a difficult question and one of the reasons why I suggested classing the cards by manufacturer and model. It would give a more accurate representation of how each exact card/manufacturer/models perform.

For instance we have specified the 2060 KO, what happens when another 2060 comes in?

[foldinghomealone]

For instance we have specified the 2060 KO, what happens when another 2060 comes in?

Currently it would show different GPU types, like:
- 2060
- 2060 super
- 2060 KO

I think this makes sense, as those three GPUs use different HW configurations and/or chips.
Making an average doesn't make too much sense because the SUPER performs much better than the others.

Maybe I don't get your point.

[Juggy]

For instance we have specified the 2060 KO, what happens when another 2060 comes in?

Currently it would show different GPU types, like:
- 2060
- 2060 super
- 2060 KO

I think this makes sense, as those three GPUs use different HW configurations and/or chips.
Making an average doesn't make too much sense because the SUPER performs much better than the others.

Maybe I don't get your point.

In the greater scheme of things a KO is just a regular 2060 and may or may not get the Turing 104 chip right so why is it classed differently to a standard 2060?

Or, are you treating that as extenuating circumstances in which case it may be a good idea to identify which ones have the TU104 and which have the TU106 chip.

My point and suggestion was to split the cards out even more granualary so people can make a decision to purchase based on an actual manufacturer/model. This may very well be too much work and maybe not worth the hassle anyway.

Just a thought

[NoMoreQuarantine]

In the greater scheme of things a KO is just a regular 2060 and may or may not get the Turing 104 chip right so why is it classed differently to a standard 2060?

Or, are you treating that as extenuating circumstances in which case it may be a good idea to identify which ones have the TU104 and which have the TU106 chip.

My point and suggestion was to split the cards out even more granualary so people can make a decision to purchase based on an actual manufacturer/model. This may very well be too much work and maybe not worth the hassle anyway.

Just a thought

EVGA has stated that all 2060 KOs will come with the TU104. They are the only manufacturer who has done this as far as I am aware. I can confirm that mine is a TU104 according to GPU-Z. That said, it may make sense to relabel it in the database as "2060 TU104" as they do on Techpowerup and leave TU106s labeled as just 2060s.

[NormalDiffusion]

Just to add a little bit more to the confusion. I got 2 more Radeon vii, so 3 in total. What I can see is just crazy! One was performing with 1/3 of the ppd. Moving her to another pc brought her to the normal value (I thought she was doa). The pc where she got 1/3 of the ppd is a e5-160 v4, so not really slow. On a gen2 16x Pci-e the radeon vii is loosing 30% of her ppd...
on the other side, a 2070super on the e5-1650 v4 has normal ppd. As well as on a gen2 16x Pci-e (xeon x5670 to be precise...).
Wondering why I was not writing and submitting data the last days? Well, I'm in hell...
And trying Linux made it even worse...