UPS Optimization - the full belt myth

[Belter]

Is a full belt gives better UPS than a partially filled one? This topic came up again in Discord and Reddit.

Measurement

Differnce is really small, for extra huge number of belts ( 100 tile long back and forth loop x 100 -> 640.400 blue belt segments) I see 8% UPS difference:

AlmostFullvsPartialBelt.png (32.85 KiB) Viewed 6621 times

Maps

Created two maps (region cloner burned my CPU badly. If someone can make a 10x bigger of this - feel free!):

- Almost full belts: https://factoriobox.1au.us/map/info/ad7 ... 15dd650aa8
- Partial: https://factoriobox.1au.us/map/info/a77 ... aa74a04389
Sry, managed to give them the same names...

Links

- fff-176
- https://www.reddit.com/r/technicalfacto ... &context=3
- https://www.reddit.com/r/technicalfacto ... ne_splits/
- Factorio-Benchmark-Powershell

[azesmbog]

Bullshit test, nothing.
Here
14493, 14493, 14286, 14085, 14085, 14085, 14085
and so
14286, 14085, 14085, 14085, 14085, 14085, 14085
if you discard the first two results (14493) - you won’t guess at all where which test is

Factoriobox results taken under ideal conditions.
Moreover, on the screen in the powershell there are slightly different numbers, the script calculates in its own way and rounds. (PS log has not yet been deleted, I can provide it for close study
So it's all within the margin of error.

[mmmPI]

What if you use only one side of the belt but use 2 belts ? same throughput of item = 1 full belt.
What if you use 2 belts filled with 50% using both side ? vs 1 fully compressed belt, same throughput of item = 1 full belt.
What if the belt is 50% filled using both side ? vs 1 belt using 1 side, same throughput of item = 1/2 belt.
What if you double the previous setup and compare the result with the first 2 ?

Because if the belt misses X(small) number of item, and the gain of UPS is X(small) , for the actual efficiency of carrying item, the purpose of the belts, the results will be difficult to gauge. If (number of item on filled belt)/(number of item on partially filled belt) =(UPS filled)/(UPS partial) then it's not conclusive on UPS. Maybe the partially filled belt is missing 8% item, maybe 16% maybe 4% which would offset the point at which one system is preferable regarding the cost of CPU time to move item around.

[Belter]

Bullshit test, nothing.
[...]
Factoriobox results taken under ideal conditions.
[...]
So it's all within the margin of error.

Let me disagree with bullshit test, sorry. I've uploaded to Factoriobox for easy sharing. Doing a test w/o mods using the PS script is the most accurate measurement I'm aware of. Actually I've found that turnign off Google drive, AV and Chrome browser gives even more consistency. If you know a better (more consistent and to answer the question) way to measure, the stage is yours.

I agree w/margin of error. If 8% comes up on my slow CPU and this amount of belts, we can close this topic I believe.

[azesmbog]

When I say ideal conditions, it means that my computer was created only for a single game (guess which one?) Even this text I am writing on another computer.
I have quite a few slow processors, I even have a 6400, though with the letter "K", I can accomplish the feat and try to run tests on it, and show an insignificant difference.
I'm more attracted to tests that even on top processors show 30 UPS / In them, belts can be neglected for sure)
So in this respect the scene is uniquely yours.

[quyxkh]

I think it might be putting things on and taking things off belts is cheaper when the lane's kept full.

[Belter]

my computer was created only for a single game (guess which one?)

Thinking...

Even region cloner was not designed for this task. I took ages to create those maps, maybe adding some space would help, don't know. Doing a belt side loading test is something I might do - it did show a diff at UPS Optimization - Red Science (editor). If you could run those FactorioBox saves on your machine would be nice just for the sake of testing

[azesmbog]

Thinking...

Bingoooooooooooo!!!

I made a mistake above.
I have an i5-6400 "not K"
Here is the difference between yours and mine:
Base frequency 3.20 GHz vs 2.70 GHz
Maximum frequency 3.6 GHz vs 3.3 GHz

Quickly tested one test:
https://factoriobox.1au.us/results/cpus ... 055c1b5f06
Processor Intel Core i5-6400 912.4 913 , 912 , 911
Processor Intel Core i5-6500 616.0 620 , 617 , 616 , 614 , 613 , 610 , 610 , 563

Should I check further or is there something to think about?))

[Belter]

https://factoriobox.1au.us/results/cpus ... 055c1b5f06
Processor Intel Core i5-6400 912.4 913 , 912 , 911
Processor Intel Core i5-6500 616.0 620 , 617 , 616 , 614 , 613 , 610 , 610 , 563

I have zero idea how can you get 1.5x UPS w/a slower(?) CPU also on Windows, please enlighten me. So I have shitty memory modules?

2x 4 GiB at 3644 MHz
vs
2x 8 GiB at 2400 MHz

CPUz-cpu.png (131.15 KiB) Viewed 6341 times

CPUz-m.png (78.09 KiB) Viewed 6341 times

[azesmbog]

Take away. 11x5 results, at the same memory frequency, on the spot. may come in handy.

If I had an i5-6500 - the results would be higher ))
In addition to playing Factorio for many years, I'm also a long-term member of hwbot.org, if that's anything to say. I have hundreds of such screenshots from CPU-Z for hundreds of processors, there are even validated ones.

[Belter]

Well, I was in OC many-many years ago. Bought this mainboard from my colleauge as he upgraded

So you say to me that my old-a.s ASrock Z170 Pro4 mainboard which
"Supports DDR4 3866+(OC) memory modules"
could be able to produce 1.5x UPS in case I go and buy a faster memory kit??

Like DDR4 KINGSTON FURY RENEGADE 3600MHZ 32GB - KF436C16RB1K2/32 (KIT 2DB) - CL16?

CPUz-mb.png (79.27 KiB) Viewed 6302 times

[azesmbog]

A little wrong.
The motherboard is good and there is a pretty big chance to succeed
The fact is that according to all Intel specifications, processors without the letter "K" are not subject to any overclocking, therefore the memory cannot be overclocked.
But many centuries ago, Supermicro engineers discovered a loophole that in motherboards based on the Z170 chipset, it is still possible to overclock "non K" processors by increasing the bus frequency, instead of 100 MHz (102 at best) to higher frequencies. My motherboard is currently running at 170 MHz with a multiplier of 27.
So, the rest of the motherboard manufacturers have released UNOFFICIAL! bios for your boards.
I have MB from ASUS, but with a lot of desire and some luck, you can find bios for other motherboards, and maybe Asrock has such a bios, they are all called "non K".
Therefore, in the general case, the order is this: first we look for a suitable BIOS, then we flash it, and only then, if necessary, we buy a good memory. But I will warn you, as always, that this is all done at your own peril and risk. The most unpleasant thing is that in case of failure, the bios will have to be programmed in the programmer.
Here everyone decides for himself.
There is another way for such motherboards - to install the next generation processor
I also had the 9700K running fine on this board, which was not officially supported by the Z170 chipsets. (my results with this processor are also very numerous on factoriobox)

https://hwbot.org/benchmark/cpu_frequen ... nterval=20
The top lines of the rating are, of course, extreme, but even without fanaticism, I think the bar of 4500-4700 MHz is quite achievable for this processor

[Belter]

first we look for a suitable BIOS, then we flash it, and only then, if necessary, we buy a good memory500-4700 MHz is quite achievable for this processor

Thank you so much! I'll dig into this.

[azesmbog]

https://winraid.level1techs.com/t/offer ... hive/36572
for Z170 Pro4 there are even two bios in the archive.
The bios is installed in the socket, this is 99% success.
I would try))

[Belter]

BIOS version is so old it is OC.. I've set BCLK 100->110 just why not. Got 620->709 UPS. I OC the CPU currenlty I know...

TY again! This is offtopic here...

lol.png (7.67 KiB) Viewed 6251 times

[disentius]

Hello Belter:)
To go back to the topic: I downloaded your saves from factoriobox. There were mods active, so I made my own.
I made a belt loop from 2k belts, and copied it 1000 times(so 1001 loops)

It is surprisingly difficult to fill a belt without some inequalities. if anyone got a good system, please share.

There seems to be a difference between full belts and not-full belts, the latter being better.... Can you check my saves and confirm/falsify?

belt ups testing-1.jpg (328.52 KiB) Viewed 6214 times

[Belter]

Hello Belter:)
[...]
There seems to be a difference between full belts and not-full belts, the latter being better.... Can you check my saves and confirm/falsify?

Hi!

1M belt segments, woah... For these saves I get the same results (after upgrading to 1.81..): not-full is faster.

full belt: 1977.5 UPS
notfull_belt: 2755.8 UPS +38% UPS

10.000 ticks, 3 runs avg.

Your results are on avg 2956 and 3948 UPS +33%.

Not sure why this difference (partial belt is filled differently, yes), but still 8% -> 38%.

At least the results are the same: difference is not significant for any real-world scenario.

edit: added avg numbers from your runs

[Belter]

OK there is a big difference between the saves: I have very long lines and only 200 lines / 200 U turns. Your saves have 2000 lines / 2000 U turns. So more U turns (10x) and 1.5x belts.

So partial belts with turns are the worst for UPS :-]

[disentius]

partial belts are faster:)
I did the same test with straight belts, Ill post it tomorrow. same results: full belts are worse for UPS...

[farcast]

Can you do a test comparing partial and full with equivalent throughput? For example, 4 100% full loops versus 5 80% full loops?