BANANA Benchmark Results Comparison

Ryzen 3900X currently on stock settings.
Corsair vengeance LPX 2*8gb 3200MHz CAS-16 (got a note on this later
Asus prime x570 pro motherboard
Nvidia GTX1070
128gb sata3 ssd and a 960gb sata3 ssd

EKWB velocity CPU water block, EKWB GPU water block, 240mm radiator and a 140mm radiator plus a D5 pump.

I am yet to settle down and dial in any cooling settings an finding I have zero need to messing with the CPU. Certainly not exceeding 65c right now with 25c ambient temp but not in love with the default fan curve

But a note that most folk don't seem to realise with memory kits. My kit is sold as a 3200MHz kit. It's actually a 2133MHz kit, but Corsair verify that it functions at the higher speed and will warranty at that speed, but motherboards will autodetect it as 2133 still, so I have forced 3200 and that's the only change to motherboard settings at this time.

 

Stock i9 9900k

 

I don't know if someone has my exact cpu, but still:

i5-7600K not overclocked

---

If you want, I'll try my best to keep the sheet in the first post up to date.

 

I did do 14 cars on the automation test track, didn't check frame rate but felt great whole time. Just, my methodology for hitting 14 cars was basically "I'm bored of this one/crashed this one, need a different one" - spawn another.

 

I managed 106.577 MBeams/s with the following hardware (no screenshot because the window closes itself as soon as it finishes)
i5-3570k @ 4.3 GHz
16 GB RAM @ 1600 MHz
Nvidia GTX 1060 6GB
Sandisk 120 GB SSD

Based on how well Ryzen is doing, I will probably be buying a 3rd gen processor soon

 

Nice setup! once i get the 3900x I may upgrade my cooling

I was able to load 16 cars and have them run in chase AI, my frames dropped from about 230fps to about 25fps, I think with 16 cars loaded the CPU has to communicate between its CCXs possibly adding some latency, I read that in a single CCX data can move about 175GBs but between both CCXs the data moves at about 30GBs I heard that the impact is relatively small for programs/games, But BeamNG is a different kind of game so I don't know what this translates into for BeamNG performance, but I was thinking of checking Gigabyte BIOS downcore control to isolate CCX performance. I think the 3000 series has already fixed that communication latency issue. I dont know weather to post that here or create a thread.

 

3000 buffs it and windows 1903 includes scheduler tweaks that help, but there's still CCX comms end of the day. Can see there's a few places in my 48 vehicle chart where it drops a bigger chunk at a time, roughly in multiples of 6, CCX's on a 3900X are 3 cores (4 with one disabled) and the windows scheduler prefers to fill a CCX before it'll start filling another, and also prioritises keeping all threads associated with a particular process in one CCX where possible.
Unfortunately I'm away this weekend, but I'd like to get more measurements and average them properly, plot a graph.

 

Nice I was hoping the scheduler would make use of single CCX where possible, A graph would be awesome to track/trace performances, I'm assuming the windows scheduler assigns vehicles for BeamNG, if not I hope the Devs implemented a vehicle scheduling system like windows build 1903. I still want to test different performance yields by way of different CCX usage scenarios for BeamNG just to confirm some hypotheses.

 

windows assigns vehicles, win32 doesnt allow setting of affinity for threads, only for processes.

 

I decided to push for 4.5 GHz with my i5-3570k. It got me 112.343 MBeams/s, but at the cost of worryingly high (1.37 volt) Vcore, so I probably won't keep it that way.

 

cmon, 1.37v is nothing for ol ivy

i've had to push my sad 4690k to 1.45v to just reach 4.6ghz (and run it that way basically 24/7 for five years)

devil's canyon is good for oc, my ass

 

Maybe, but I was hitting 100 C running prime 95 even with an H100i AIO cooler, which was concerning

 

well, yeah, so do I, in the summer, i can last for a few minutes in the winter, with an h100i

prime95 isn't a realistic workload for anything other than prime95 though, i get ~100 watts consumption with "typical" 100% loads, ~175 under prime95 small ffts

i mean i do have mine delidded/liquid metalled/lapped, but that was just to see how high i could really take it

and the answer was none higher than i already was

 

I messed around with some more settings and things seem mostly stable. I guess I'll keep it at 4.5 as long as things don't get too funky.

 

it just really busts my balls that i had to work so much to get 4.6ghz and these 14nm+++++++++ bastards get 4.7ghz all core stock ;-;

 

Prime 95 small FFTs on a 3900X I was reaching 115W as measured by HWMonitor and sitting just shy of 70c

 

Not to bad for a gaming laptop.


CPU Type HexaCore Intel Core i7-8086K, 4500 MHz (45 x 100)
Motherboard Name Clevo P7xxTM1
Motherboard Chipset Intel Kaby Point Z370, Intel Coffee Lake-S
System Memory 32714 MB (DDR4 SDRAM)
DIMM1: G Skill RipjawsV F4-2400C16-16GRS 16 GB DDR4-2400 DDR4 SDRAM
DIMM3: G Skill RipjawsV F4-2400C16-16GRS 16 GB DDR4-2400 DDR4 SDRAM

Display:
Video Adapter GeForce GTX 1080 (8 GB)
Resolution: 3840 × 2160
Monitor AU Optronics B173ZAN01.0 [17.3" LCD]

 

To be fair, that is literally a desktop processor in there, those clevos are cool machines

 

Done, you're welcome (pretty heavy, watch out)
Banana Bench Results - NOW WITH MORE THAN 4 CORES!
You can filter everything, just click the small funnel:

And then click "Create new temporary filter view".
You may notice we now have 4 sheets:

• "Old Banana Bench" contains the old banana results (Pre-Race until 0.5.x versions);
• "New Banana Bench (Dynamic Collisions ON)" contains all the results from 0.5.x to 0.9 basically;
• "New Banana Bench (Dynamic Collisions ON) Ver.0.9+" it's self explanatory: the results from 0.9 until now; I advise you that some of those scores are updated with more recent ones, others are untouched from 0.9.x/0.10.x, so results may vary;
• "New Banana Bench (Dynamic Collisions ON) Ver.0.18+" contains the benchmarks after the 0.18 update in which the test vehicle (the pickup truck) got updated;
• "New Banana Bench (Dynamic Collisions ON) Ver.0.26+" contains the benchmarks after the 0.26 update as it brought massive performance improvements across the board;
• "New Banana Bench (Dynamic Collisions ON) Ver.0.28+" contains the benchmarks after the 0.28 update as the test vehicle (D-Series) got updated;

Notes:
CPUs with "N/A" in "J Column" are CPUs that:

• didn't meet the minumum requirements (basically 0 cars at 100% realtime);
• didn't have the bananabench screen;

For CPUs without a specified clockspeed, I've put the maximum speed of the turbo boost.
Most of the newer results have as a maximum of spawnable cars 20, that's a Banana Bench limitation, not my fault; capped tests are highlighted with a different background colour.
I'd like to remember that if you want your score to be added to the sheet, please follow these guidelines:

• Your test must be done with the default vehicle (pickup) as in 2014 it started with that, you can post results with different vehicles but I won't count them;
• You can use both the original Banana Bench or the in-game one, just attach a picture of the test at your answer;
• For sake of simplicity, write the speed of your CPU during the test (if it turboes up) or if it's OCed, otherwise I have to put the manufacturer's maximum turbo frequency;

---

There may be errors, if you spot one, tell me!
Sources:
https://ark.intel.com/content/www/us/en/ark.html
https://en.wikipedia.org/wiki/Comparison_of_AMD_processors
http://www.cpu-world.com/index.html

Spoiler: Changelog

14/07/2019: Fixed spreadsheet link
14/07/2019: Removed Spreadsheet link until Tomorrow, sorry for the inconvenience
15/07/2019: Spreadsheet is back online
13/01/2020: Fixed order of sheets.
13/12/2020: Revamped spreadsheet design.
12/04/2021: Added "Useful and interesting data" page.
08/10/2022: Updated to latest Bananabench.
10/11/2023: Updated to latest Bananabench.