Discussion PPL - Performance Per Liter - Round 3 - Compute Edition...?!?

thelaughingman

Master of Cramming
Jul 14, 2018
500
491
Last try - hitting 25k
(6,920 x 7,118 ) / (7.28^1.5) / 100 = 25,077
  • Unshackled the CPU, PBO enabled, no longer in Eco mode. As you can see the BR clone barely keeping up with it as temps creeped towards 95C
  • Lost 150 points on the CPU as my ambient temps has changed from 20C to 32C within 2 weeks haha

 

theGryphon

Airflow Optimizer
Jun 15, 2015
298
236
So you dropped bricks from volume? I don't think that's fair, although it gives me a boost 😀
 

Aux

Cable-Tie Ninja
Dec 5, 2018
160
126
So you dropped bricks from volume? I don't think that's fair, although it gives me a boost 😀
Ooops . . Apologies, a mistake on my part. I forgot to modify the calculation. If you notice, your position is where it would be with the correct calculation. I was trying to be clever and show all the volumes . .. Just correcting now . ..
 

rfarmer

Shrink Way Wielder
Silver Supporter
Jul 7, 2017
1,929
1,906
Well I upgraded to a RTX 2080 Super and removed my tophat so this will be about as good as my PC will perform until I upgrade to Ryzen 4.

Case: Louqe Ghost S1 8.70l
ghostbs2back.jpg

Cinebenchr20.jpg

timespy.jpg

((3885*11740)/(8.70^1.5))/100 = 17773.7980424
 
Last edited:

PLLovervoltage

Efficiency Noob
May 26, 2020
7
10
Love this PPL metric! Got me to sign up to this site.

Can I suggest a slight change to how the calculations are done? Currently, the performance scores are a simple multiplication which will be biased based towards either cpu or gpu depending on the scale. Using the geometric mean instead will normalize the range of the values so that cpu and gpu performance is equally weighted regardless of the scale used.

Current: ((CPU*GPU)/(Volume^1.5))/100
Geometric mean: ((CPU*GPU)^0.5/(Volume^1.5))

You won't have to divide by 100 also. I'm guessing the cinebench and time spy was chosen as they had somewhat similar number ranges. Using the geometric mean, will give you a better unbiased weighting between cpu and gpu performance. It'll also give you the freedom to use any other benchmark with widely different number ranges without bias towards either the cpu or gpu.
 
  • Like
Reactions: thelaughingman

theGryphon

Airflow Optimizer
Jun 15, 2015
298
236
Love this PPL metric! Got me to sign up to this site.

Can I suggest a slight change to how the calculations are done? Currently, the performance scores are a simple multiplication which will be biased based towards either cpu or gpu depending on the scale. Using the geometric mean instead will normalize the range of the values so that cpu and gpu performance is equally weighted regardless of the scale used.

Current: ((CPU*GPU)/(Volume^1.5))/100
Geometric mean: ((CPU*GPU)^0.5/(Volume^1.5))

You won't have to divide by 100 also. I'm guessing the cinebench and time spy was chosen as they had somewhat similar number ranges. Using the geometric mean, will give you a better unbiased weighting between cpu and gpu performance. It'll also give you the freedom to use any other benchmark with widely different number ranges without bias towards either the cpu or gpu.
Multiplication already balances CPU and GPU measures. No need to take a square-root there. Not saying it's wrong but it brings nothing to the table.

Plus, going with your formulation, volume becomes much much more important. Your ranking is equivalent to

((CPU*GPU)/(Volume^3))/100

Compared to the current ranking, it makes the volume dominate. Is this desirable? I'm not sure. I mean there can be an argument to make volume more prominent than it is now but even then, I would recommend going for a power of 2 instead of 3. So it'd be like:

((CPU*GPU)/(Volume^2))/100

Or, if we go with your approach, it can be

((CPU*GPU)^0.5/(Volume))

They're equivalent, ranking-wise...


EDIT: I believe it is fine as it is now. Increasing the power of the Volume takes away from the definition of SFF, if by SFF we're talking about sub-20 liters. Another argument can be even to reduce the power of Volume from 1.5... Volume in itself is a cubic measure (mm^3, etc.) so why further increase its exponent?


EDIT 2: Now that I started thinking about this, I gave it a good thought... I really believe taking Volume to a higher exponent than 1 takes away from the PPL measure: it must be (CPU*GPU/Volume) as it was originally. Any power other than 1 is simply arbitrary and it rewards micro-FF systems for no reason. PPL = performance per liter after all ;)

CPU and GPU performance benchmarks can evolve over time, of course, and it needs to be debated whether the current benchmarks (CB20 and TimeSpyGFX) are the best ones... One idea can be to use two CPU benchmarks and two GPU benchmarks (and possibly take a geometric mean at that point).
 
Last edited:

DrHudacris

King of Cable Management
Jul 20, 2019
646
1,249
Your ranking is equivalent to

((CPU*GPU)/(Volume^3))/100
I think you are incorrect here. The suggested formula and yours are not equivalent.

EDIT: I believe it is fine as it is now. Increasing the power of the Volume takes away from the definition of SFF, if by SFF we're talking about sub-20 liters. Another argument can be even to reduce the power of Volume from 1.5... Volume in itself is a cubic measure (mm^3, etc.) so why further increase its exponent?
This is also incorrect. Yes, volume is calculated as a cubic of distance, but it does not matter as this is thread is performance per liter, not performance per length.
 

theGryphon

Airflow Optimizer
Jun 15, 2015
298
236
I think you are incorrect here. The suggested formula and yours are not equivalent.
I think you should read more carefully. I didn't say the formulas are equal. I said, and I quote "rankings are equivalent". That means both formulas would produce the same "ranking" of systems.

This is also incorrect. Yes, volume is calculated as a cubic of distance, but it does not matter as this is thread is performance per liter, not performance per length.
I'm not sure what you think is incorrect in what I wrote... I said to use directly Volume (measured in liters, as you know I hope?) and not any power of it. [Edit: My second edit in my previous post was not in for some reason. I fixed that now. There I was offering to use Volume directly]

And I hope you understand what a liter is?
1 Liter = 1000000 mm^3
So, yeah, liter is not a distance measure, it is a measure of volume.
 
Last edited:
  • Sad
Reactions: DrHudacris

PLLovervoltage

Efficiency Noob
May 26, 2020
7
10
Yes. I was wrong by ignoring the power on volume thinking it didn't matter but it does have a very major impact.



Geometric mean for 2 terms = (a * b)^0.5

Current: CPU*GPU / Volume ^1.5
Correct: Geometric Mean / Volume
Incorrect: Geometric Mean / Volume ^ 1.5
Simple: CPU * GPU / Volume ^ 2


Ranking wise they are nowhere equivalent. I only looked at the top 5 but there's a very large difference in how systems are ranked based on the equation used. I was wrong to ignore the power on volume. Like you said, we're measuring performance per liter and putting a power other than 1 on volume makes a drastic difference. My initial suggestion was wrong by ignoring the power on volume. The correct method is to use the geometric over volume with a power of one. Or as you said, the simpler and equivalent equation would be cpu*gpu over volume with a power of 2. The equation in this format may seem like we're giving volume more weight but mathematically speaking.. what's actually happening is we're getting the geometric mean of the cpu and gpu benchmarks.

In conclusion, what we want to measure is performance per liter. If we want to respect that, we have to use either the equations (Geometric Mean/Volume) or (CPU*GPU/Volume^2)

Regardless, I think my message went across. (Geometric Mean/Volume) or the mathematically equivalent (CPU*GPU/Volume^2) should be implemented to correctly represent performance per liter.

Referring to @theGryphon's suggestion, getting the geometric mean of 2 cpu and 2 gpu benchmarks over volume of a power of one would be a very nice and accurate measurement of performance per liter. I'm pretty much a noob to the sff community currently with a mid-end node 202 system(hoping to upgrade and downsize soon) but would there be significant difference between 2 different cpu/gpu benchmarks? I also imagine what benchmarks that are used (such as single/multi core cpu performance) would be a pretty contentious topic. I will concede to whatever the more experienced sffpc community chooses. However, I personally very strongly advise that in order to correctly measure PPL, the equations I proposed be implemented.
 
Last edited by a moderator: