CPU Performance Formula

Hey Sup Forums, my friend is building a computer, but he doesn't understand much about CPUs.
He wants to find himself the perfect CPU for his budget, so I made the following formula (pic related):
Performance Points = [(C * 2) + T] * TF

Where:
> C = Number of Cores
> T = Number of Threads
> TF = Max Turbo Frequency

He wants 60fps 1080p gaming, so I told him anything close to 50 points is fine for the CPU.
I know it's a really rough way to preview the performance, but since AMD and Intel have really close IPC performance nowadays, I think it works for nomies.

What do you think Sup Forums? How would you improve it?

Other urls found in this thread:

en.wikipedia.org/wiki/Megahertz_myth
sites.utexas.edu/jdm4372/2016/11/05/intel-discloses-vectorsimd-instructions-for-future-processors/
twitter.com/NSFWRedditGif

>since AMD and Intel have really close IPC performance nowadays, I think it works for nomies.


AMD IPC is still ~5% lower.

Forgot to mention, but it obviously doesn't work on AMD FX chips

> ~5% lower
close enough

bump

I changed the 50 points to 30 for 1080p gaming.

op if you are just gaming on 1080p just go with a fucking g4620 and a 1070 literally will be able to play any game at high refresh rates

OP here, updated with the final version.
Now it's more realistic when compared to benchmarks.
> pic related

yeah, that's why I said 30FPP should be enough for 1080p 60fps.

Can you do that with cars?

Performance = [(Wheels x wheelsize) + nitro] * gearbox

No? Why not?

I'm not into cars, but the formula is just measuring a single part of the whole computer. I'm positive that you could measure the performance of some part of the car's engine with a formula as well.

It seens to align well with Geekbench as well, I should look into games tho

>All cores are equal

Look into the updated formula:

When discussing CPU value, make sure to factor in motherboard price as well. Coffee lake chips need rather pricey z370 motherboards, while for most people a cheap B350 board is all they need for a Ryzen build.

pic related is the i3 8350K vs R5 1500X

The formula measures pure performance alone, if you want to make a price/benefit formula, just make a mobo constant list and multiply by the FPP I guess.

Doesn't this heavily weigh in favor of core count?

Many games rely on single core performance.

A 12 core 24 thread Xeon from 2013/14 will max out around 3.5ghz single core turbo frequency and has an FPP around 168
A 4 core 8 thread 7700k will max out at 4.5Ghz and has an FPP of around 86

So even though the 7700k is VASTLY superior for gaming, the several generation old xeon looks very impressive.

Is that pre or post patch?

The formula previews that a OCd i3 8350K will beat a OCd 1500X, which aligns well with benchmarks.

Of course it doesn't take in count mobo prices, just a fun fact.

Max turbo frequency shouldn't have the weight it does in the calculation. Intel's crappy coolers can't maintain boost clocks for shit. If you're gonna factor boost clock into the score, factor in the price of the cooler to maintain it.

The formula gives an overall performance view, it's not calibrated for gaming.

I imagine a gaming formula would weight the frequency a lot more than threads/core count.

Now the final battle:
i7 8700K 5GHz vs R7 1800X 4.1GHz

i7 8700K
> FPP: 150.0

R7 1800X
> 150.9

Well, that's a match for me (of course the i7 will get better gaming performance, but overall performance will be similar, the 1800X will probably still beat it on productivity)

I don't have words to tell you how stupid you are for thinking this.

> numbers don't define a car's performance

b8

kek most likely pre, those sore intlets are still in denial over Specdown

Now, in all honesty, Intel can't compete with the R7 1700 AT ALL (pic related)

8c/16t for $290, not even in the wildest dreams Intel could imagine AMD would accomplish this. The 1700 just kills anything Intel has at that price range, and the formula shows that as well.

Heres the deal though, are you willing to shell out 30 dollars more for a k proccesor that has just 5% more ipc?

Opinions on pic related?

> changed from FPP to OPS (Overall Performance Score)

I think you really need to factor IPC of each chip dude. Its kindof important regarding raw power and clock speed.

See: GC is basically IPC

>stupid compsci faggots try to into science
The units on your formula don't make sense, idiot.

why not?

Considering performance per chip can easily fluctuate by 5% depending on everything from fabrication inconsistencies to individual system components to operating environment, I think it's more than fair to say that a 5% margin is "really close", and that you're being overly semantic for trying to highlight - or worse yet, correct - the difference.

OP here, here's a new image.

he's right tho, IPC is important. The new formula fixes that, take a look at pic related.

Is the Meltdown performance hit included in tge equation? I propose we reduce 0.30 for all intel cpus

If you can get the data you could do something like this:
>collect benchmarks on cpus
>collect as much attributes about the cpus as possible
>learn a model to predict benchmark score

Thought about it a few weeks ago that you could technically model the benchmark of a build before buying it with enough data

not necessary, Intel is already geting BTFO'd the way it is.

Hold them accountable to their actions.
Reduce the score please

>IPC
>IPC
>IPC of each chip dude. Its kindof important regarding raw power and clock speed.
>IPC is important
IPC is not a very meaningful figure because it depends on _which_ instructions you're executing.
Two CPUs could be identical in muh games but if one has AES-NI, for instance, and the other one doesn't, there will be a tremendous difference in IPC if you measure encryption.
One CPU might be better at branch prediction than the other, etc..
Real world benchmarks is the only meaningful way to compare CPUs.

What multiply C * 2? It should be this:

(C + T) * TF

If something has 4 cores, it should have 4 cores. Multiplying it by 2 seems kind of arbitrary.

performance points:
amd: 0
intel: (inf)

>some workloads could be effected up to 30% with the meltdown/spectre patch
>90% of use cases will see no performance decrease at all
Yeah, fuck it. Might as well just lower their scores by 30% because fuck using logic...right?

Turbo is shit though, it lasts for micro seconds

Cores > Threads

If C had the same weight as T, that would mean a 6c/6t CPU would match 4c/8t CPUs, but that's not the case, the 6c/6t actually outperforms it a bit.

See 8th gen i5's vs 4c/8t R5's

Maybe your turbo from 2011.

Modern turbo boost is sustained forever assuming your cooling can keep up.

Further, with turbo boost 3.0 there is a lot of granularity to turbo boost, single core vs all core turbo, etc.

Not if you have decent cooling

X79 system:
{[(6x2)+6]x3.9}x1.0x0.90(post Spectre mod and Meltdown patch)
=63.18
Ivy Bridge needs a larger gap from Haswell after the Meltdown patch because it lacks features(optimization) to mitigate the slowing down of the system after the patch.
I recommend those who are on Haswell and newer with the Windows Update and µcode update to deduct 0.05(Windows Update only -0.02), and those who are on Ivy Bridge with the Windows Update and µcode update to deduct 0.1(Update only -0.07). Ivy Bridge takes a bigger hit due to the lack of optimization features that can mitigate the slow-down of the system due to the Meltdown update.

Posting the CPU list I made so far, it seems to works quite accurately desu

user something more important than cores since you already are counting the threads would be instructions per clock.
Also this simple formula would even work with processors who don't have multithreading since they most likely have a higher instructions per clock count to begin with.

ex: performance = total thread amount * (instructions per clock * max clock all cores can reach together)

The reason I said all cores can reach together is that some turbo technologies will vary for example 3.6ghz on half cores or 3.0 on all and so on you can check this part on the manufacturer database.

> IPC
see:

Autist

A few extreme chips

I had an issue with OP using max turbo as a baseline, but this actually goes well with single-threaded performance figures at the expense of skewing full multi-thread utilization.

One does not simply make these comparisons.
Here's how you're wrong.
en.wikipedia.org/wiki/Megahertz_myth

Most people use IPC as a conglomerate factor taken from multiple single threaded tests as you'll find in reviews.

>Real world benchmarks is the only meaningful way to compare CPUs
But the formula just werks. It's all about cores/threads/clock user, it's the same arch.

Check out the new formula:

OP included a rough by-the-generation IPC factor and showed that his formula generally compares well to benchmark scores.

Solid work OP, always good to see simple calculations make sense. This is as good as the power draw formula for overclocking.

Same arch but with different instruction set extensions. For instance, does AMD support this bad boy?

sites.utexas.edu/jdm4372/2016/11/05/intel-discloses-vectorsimd-instructions-for-future-processors/

Did you know that hardly 3% of programs effectively utilize 7-year-old AVX, let alone AVX2?
Did you know this thread is about general processor applications, not niche HPC vector code?

Which, granted, is an AVX-512 instruction. But I'm sure you'll find a difference between the instructions supported by AMD and Intel processors.

>DGEMM is niche
Since when? :(

Oh silly OP and his AMD Shilling Formula.
Do one for GPU's too and remember #ofshaders > everything else!

>what is IPC
What a shit formula

see:

Not with high speed low latency ram

>spending even MORE money making up for inadequate architecture.

I'm personally holding my breath for a solid showing from AMDs Zen 2 or 3.

Same with Intel and their 10-7nm products.

Competition is good.

>[(C * 2) + T] * TF
What do you hope to accomplish by this midnight calculation? Are you expecting the results of this formula to match up with any real-world, practical performance numbers whatsoever? Have you tried whether it does?

Making better RAM is not "making up for inadequate architecture." You can just keep improving CPUs, you have to improve memory as well, which hasn't been happening.

Should do an OPS per dollar calculation as well

Why don't you read the thread and find out?

this doesnt really work that well, more than 4 cores do not give extra performance for the most part. single thread perf is the most important for gaymes

You shut every other program off when playing games?

It works user
> pic related

except for "muh gaymen"

>how would you improve it
I'd probably add in terms for each type of cache with inverse significance (L1 being least significant Ln being most significant)
Because realistically you only care about how well your computer performs for shit applications and those are largely cache-unaware.
Instruction set support is also relevant but just like with caches they're less well used the worse the software is. Because automatic vectorization isn't done well.

And then you should consider adding a couple tuning factors in there, download a complete list of CPU benchmarks and adjust the factors to fit the stats.

btw, I calibrated the GC to favor Ryzen a little bit less, now it's even closer to real world benchmarks/gaming.

> cache
that would make it too complex

Thanks Doc.