CPU Performance Formula

Alright Sup Forums, I came up with a simple way to calculate a CPU's performance by just getting a few important data, such as:
> Number of Cores
> Number of Threads
> Max Turbo Boost
It works with Intel and AMD.
The formula is:
OPS = { [(C × 2) + T] × MTF } × GC

Where:
> OPS = Overall Performance Score
> C = Number of Cores
> T = Number of Threads
> GC = Generation Constant (see pic related)
I didn't think it would work so well, it's obviously not 100% precise, but it comes really close. It has a margin or error of just ~5%.

Let me show an example of it's precision:
> R5 1600: 93.7 OPS (104.1 OPS if OC'd to 4GHz)
> i5 8400: 90.0 OPS
Using the formula the R5 1600 stock is 4% faster than the i5 8400, and if OC’d, it’s 13.5% faster.
Now let's check some benchmarks:

Geekbench (Multicore):
> R5 1600 (OC): ~25500
> i5: ~22500
Here, the 1600 is 11.7% faster than the i5, which is well within our 5% error margin.

Cinebench R15 (Multi-Thread):
> R5 1600: 1134
> i5 8400: 945
That gives the R5 1600 a 17% advantage, again, really close to the 13.5% previewed by the formula, and within the 5% error margin.

… cont

Other urls found in this thread:

cpubenchmark.net/high_end_cpus.html
twitter.com/SFWRedditImages

cont…

> but what is GC? How did you calculate it?
Fist, IT IS NOT IPC. The GC, or Generation Constant, is a value that makes the final result more realistic when compared to real world results. You can use your own GC if you want. My GC list is based on AMD’s own Ryzen benchmark comparisons with Intel, which includes:
* Video Encoding
* Secure Data Encryption
* Content Creation
* Game Streaming
* 1080p Gaming
I personally believe those benchmarks are fair, but if you believe there’s a bias towards AMD, you can use your own GC values.
> But how did you calculate those values?
Look at pic related, it’s AMD’s own comparison between the R3 1200 and the i3 7100. If you make an average of the results, you will get a 7.4% overall advantage for the R3 1200.
Now, let’s use the formula WITHOUT the GC value (BPS stands for Brute Performance Score):
BPS = [(C × 2) + T] × MTF

The Results are:
>R3 1200: 40.8 BPS
>i3 7100: 31.2 BPS
That gives a 23.5% advantage for the 1200, which is WAY above AMD’s own benchmarks. That’s the reason there’s the need of a GC. Now let’s see the results with the GC applied:
>R3 1200: 42.8 OPS
>i3 7100: 38.22 OPS
That gives a 10.4% advantage for the 1200, which is much closer to the 7.4% we’re aiming to, and inside the 5% error margin. That’s also a result that’s close to most benchmarks out there.

You've just made a formula that looks about right because you have introduced arbitrary values.

This

Here's a list of CPU's and their respective OPC

Not arbitrary, the value introduced is based on real world benchmarks.

>my formula predicts well when i include a number based off the parameter i am predicting

Best CPUs based on performance/price, according to the formula:
> Pentium G4620
> i3 8100
> R5 1600
> R7 1700
> i7 8700K
> TR 1920X
> TR 1950X

It works well with most benchmarks tbqh. Benchmarks are the only way we currently have to measure a CPU's performance.

>this retard doesn't even know that (a*b)*c is just a*b*c and is trying to lecture us about his formulas
kek

>"Benchmarks are the only way we currently have to measure a CPU's performance."

I wonder why that is user. Perhaps... maybe... it's because cores and threads are not sole factors in performance.

>(a*b)*c
But it's (a*b) + c

Talking about you multiplying by GC by putting the other things in parenthesis when it's obviously not needed
also

Literal brainlet can't even read his own formula

Also, this

>it's because cores and threads are not sole factors in performance
Well, when we are comparing between different architectures, that's the price we pay. If we're comparing between Intel's 8th gen CPU's alone, we shouldn't need to introduce any value and it would work very accurately.

Seriously user, you need to make up your mind - is it OPC or OPS?

No, it would not work accurately at all.

If I had a quad core 14nm processor and a quad core 7nm processor, this would have a large impact on performance.

Not to mention differences in architectures, instruction sets, etc.

> Intel's 8th gen
> 7nm and 14nm

I'm talking about comparing between the same generation user.

OPS. sorry, that was a fuck up.

OPS = [ C × 2 + T ] × MTF × GC

Happy now?

What's the point? You're scaling it with a variable determined from benchmarks, then comparing it to benchmarks. I don't see what this formula is supposed to accomplish.

Amd slide thread using cherrypicked benches

>C x 2
>not 2C

It's supposed to save you a lot of time when researching and comparing CPUs.

Also could be used to predict the performance of unreleased CPUs by speculating it's GC, and easily compare against any other CPU.

Fixed.
OPS = [ 2C + T ] × MTF × GC

>No, it would not work accurately at all
wut?

The 'generation constant' that is actually a variable and not a constant... is obtained, how?


Your equation accounts for single thread performance(based on max clock speed), per generation. That's all it does. It then uses benchmarks to order them.
>a single thread at a lower clock speed, from the same generation, can outperform a faster thread of the 'same generation'
>because we all know the new RYZEN2 is the same generation, just tweaked?
>how do we get that 'not constant' again
>Do we have to make up a new "generation" when they are released?

SIMPLE WAY TO CALCULATE CPU'S PERFORMANCE
> cpubenchmark.net/high_end_cpus.html

>how do we get that 'not constant' again
see: >Do we have to make up a new "generation" when they are released?
every new generation will need a new and higher GC value, like Ryzen+ and Intel's 9th gen.

> SIMPLE WAY TO CALCULATE CPU'S PERFORMANCE
That's the whole point of the formula, instead of having to access some jews website, you can just use a simple formula on any CPU.

This is a shill thread, it's getting spammed now.

Is it a AMD or Intel shill thread tho?

Just updated formula and variable names.

bump

>tfw 2nd gen

probably 0.8 idk

bump

You have to be 18 to post here.
That formula is retarded, one already exists, and you don't know what the fuck precision or accuracy means, nevermind how a goddamn cpu works at even a basic level.

Posters like this ought to be publicly doxxed for being retarded

Also, you use apple so tell Daddy to use their money and buy you a fucking education

the formula just works

> one already exists
which one?

Cinebench score :^)

> memebench

I have a new formula for this.

PPS = CC (1000 -1000)

Where
PPS = Processor penis score
CC = Chad Constant

You can obtain the CC for the CPU by looking up the Geek bench score. This is actually 100% accurate for any CPU!

Oops, I'm missing a +

It is PPS = CC +(1000-1000)

Sorry, but OP's formula works offline.

Awesome. Tell me what the 8 Core Zen 2 will bench then.

We can speculate it's GV (Generation Variable) and it's BF (Boost Frequency)

For the 2800X, I imagine something like 1.35 and up to 4.6GHz.
That's ~198.7 OPS, or 40% faster than a 1800X.

Having that said, I imagine 40% is too much and maybe I'm expecting too much from Zen 2, so it probably won't boost this much, or have this much GV gain.

If AMD manages to get a 15% bump on Zen 2, then the 2800X specs will be:
> 8c/16t
> 4.4GHz Boost
> 1.15 GV

Looking at this it seems like quite an accurate representation of how the the advances in CPUs have been over the years.

rebbit

single thread perf is also extremely important though.

OP your formula is not good because it's obvious you have derived it from benchmark results, rather than came up with the formula that will correctly predict benchmark results.

The 'C * 2' part is arbitrary and meaningless. Then you come up with a 'constant' (again, arbitrary) so that the formula can fit the results.

But, you're not the first to do this and there's actually a precedent in the history of science. When astronomers looked at the orbit of Uranus they predicted its motion using Newton's laws of motion and gravitation. However, they found that its orbit does not match the prediction, there was a perturbation. Some astronomers suggested adding an arbitrary constant to the calculation that would make the equations fit the observation. But that was not a good solution to the problem. Then, an unknown planet was theorised to exist and the existence of the planet would explain the anomaly in Uranus' orbit. That planet, of course, was Neptune and it was later observed.

So you see, you might not be as dumb as people here think, just misguided. That's why study of history is required so we do not repeat the same mistakes.

If you really want to come up with some formula that can measure CPU performance from easily-knowable values you should study CPUs more, look into the design of CPUs and try to figure out what qualities of the CPU are responsible for its performance. I obviously don't want to discourage you if you're interested in CPU design you should pursue it. Everyone has to start somewhere. Just keep at it. :)

>Using the formula the R5 1600 stock is 4% faster than the i5 8400, and if OC’d, it’s 13.5% faster.
>Here, the 1600 is 11.7% faster than the i5, which is well within our 5% error margin.
5% of what?
5% of 104% is 5.2%, less than half of 11.7%.
In the best case for you, 4+5 = 9%, still 20% off, and this sum doesn't even make sense in statistics.

>We can speculate
So in order to assert, you'd have to check a benchmark. Works great...

What about FLOPS? Could be a starting point.

FLOPS is only useful if doing lots of floating point calculations. MIPS would be a better metric but even that can change based on what type of benchmark you're running.

>Have to check a benchmark to obtain a variable
>Percent difference in scores between processors ideally matches the benchmarks
Okay? Just check the benchmark score then.

Retarded
Doesn't take into account multicore turbo
"Generational constant" is a retard's way of referring to IPC but there are different instruction sets

GC is not IPC, it’s just an average value of several benchmarks

OP here. Thanks for this post user! I agree with many things!

> The 'C * 2' part is arbitrary and meaningless
Arbitrary? maybe, but meaningless? no. We all know core count is more important than thread count. That's the reason a 4c/4t core i5 is more powerful than a 2c/4t i3, also that's the reason a 6c/6t is faster than a 4c/8t.
I don't know if you noticed, but the following formula works between the same architecture/generation:
BPS = ( 2C + T ) × MTF

The BPS (or Brute Performance Score) is very accurate between the same architecture, and that's without the GC, or the "arbitrary constant". Just knowing the number of Cores, Threads and Boost Frequency, we can predict one CPU's performance against another one of the same generation/architecture, and that's obvious! Since they're basically the same CPU with different parts disabled/enabled.

The problem comes when comparing between different generations/architectures, like 4th gen vs 8th gen or Ryzen vs Intel. And that's the reason we need the "arbitrary constant". Let's be honest, the concept of IPC itself is arbitrary, since it just works on some workloads. So an IPC constant is not enough, that's why I came up with the "arbitrary value", which consists of an average of several workloads.
That arbitrary value allows us to compare between different architectures, and there will never be a formula that does that without some kind of arbitrary value.

So if you want to compare two CPU's, which one would you rather do?
> Use a simple formula on both CPU's
> Look into 5 different benchmarks for each CPU

What if the CPU is new and there's no benchmarks available yet? The formula is handy...

So you want a formula that predicts the future? How could one possibly know how much a new CPU will improve? It could be 10%, it could be 50%, we don't know!

>Doesn't take into account multicore turbo
turbo is cancer, fuck you.

So what's the point if you can't predict the future, and you have to derivate the generation factor from benchmarks?

see:

>Doesn't take into account multicore turbo
> What is the 5% error margin?

That's fucking retarded, congrats.

>this thread

elaborate.

Can't you read the thread?

So what real-world data have you compared this against and how good is the correlation over all the processor families you claim to support? I mean, surely, this isn't just something you've thought up in your armchair with no data to back it up, right?

rtft

Gimme the score for an i7 4770k at 4.1 GHz.

see:

There is literally no correlation data in the thread, though. It's all armchair calculations.

Are you saying those two numbers are the entirety of your data? Two numbers for four variables?

71.17 OPS

1.43 MD/h in my arbitrary formula that measures the time needed to render my micro dick in a accurate model with 50 millions polygons.

the GC is the result of an average of several benchmarks. Every other data comes from the specs of the CPU itself.

>Every other data comes from the specs of the CPU itself.
Except the actual formula, which is a pure heuristic. Where's your correlation data, proving that it is remotely accurate?
>the GC is the result of an average of several benchmarks
You mean those quoted benchmarks for no more than two different processors?

> Where's your correlation data, proving that it is remotely accurate?
See pic related. Let's use the BPS formula - which only works between the same generation of CPUs - for the i3 8100 and the i5 8400:
> i3 8100: 43.2 BPS
> i5 8400: 72.0 BPS
That makes the i5 40.3% faster than the i3, according to the formula, and that's without the controversial GV value.

Now, let's see if it's "remotely accurate", according to 3 different benchmarks:
Geekbench 4 (Multi)
> i3 8100: ~15300
> i5 8400: ~ 22700
That's a 33.2% advantage for the i5. A 7% difference in just one benchmark is pretty accurate, let's be honest.
Cinebench R15 (Multi)
> i5 8400: 945
> i3 8100: 568
That's a 39.9% advantage for the i5. Practically identical to what the formula previewed.
3D Mark TimeSpy (CPU - Physics)
> i5 8400 - 5472
> i3 8100 - 3750
That's a 31.4% advantage for the i5, again, just a 8.4% difference for this specific bench.

Those variations are natural, because every different benchmark performs differently. But let's check the average precision between these 3 benchmarks:
(33,2 + 39,9 + 31,4)/3 = 34.8%

So, the formula previewed a 40.3% advantage for the i5, and the benchmarks gave us 34.8%. That's a 5,5% difference, pretty close to the ~5% margin of error I previewed (and that's with just 3 benchmarks)

bump

bump

I too, was 16 once.

Your math is all over the place. How are you even arriving at those values? 72 / 43.2 = 1.666..., which is a 67% performance advantage, not 40.3%. The only thing I can see that would get close to 40.3% is 1 - (43.2 / 72), but not only can I not see why in the world you would calculate it that way, but also that is exactly 40%, not 40.3%.

The same thing is true for the other numbers, but the true numbers are 48%, 66% and 46%, respectively. 20% out is a pretty big difference.

More importantly, however, this is just three benchmarks. For four variables. You wouldn't even get any precision in *one* variable with that small sample size, but using less samples than your number of variables is beyond uneducated.
>margin of error
You don't understand this term. The margin of error isn't something you just postulate, it's something you measure by comparing your model against real data. With a functional sample size.
>the average precision
You're just embarrassing yourself. You can get any figures you want to selectively averaging some subset of benchmarks together.

Also, before you even start thinking about presenting results, the least you do is to apply your model to benchmarks not part of the sample size and see how well it fits. If the model has no predictive power, it's as good as useless.

>hahahaha I created a formula that lets you measure the performance of a CPU without using a benchmark
>oh wait, you do need to use a benchmark

>20% out is a pretty big difference.
That's actually 13,6% if you're calculating that way.
> More importantly, however, this is just three benchmarks. For four variables
Those three benchmarks are just examples. I've made many more on this thread alone.
> If the model has no predictive power, it's as good as useless
But it has. If intel releases right now the 8th gen Pentium, I will be able to predict it's performance way before any benchmark is out, just by knowing it's specs.

>oh wait, you do need to use a benchmark
No, you don't. You just need a benchmark if you want to verify if it works (it does)

Yes you do. See GC. Anyway, that's a terrible way of measuring things, since threads and clocks can't define performance alone.

This thread is still alive? OP is just a retard that think making a retarded formula that does nothing more than a rule of three makes him hot shit.