BANANA Benchmark Results Comparison

@FLAXMS
So I will have a more detailed post when i'm back at home on my desktop but I encountered the same crashing BS with banana bench with my new system I built.

i9 14900K
32GB DDR5 6400@6800
RTX 4090

I never paid much attention, but I ran BB out of the box with no changes other than no power limits, and DDR5 ram @6800 and XMP enabled. I scored 985.....which based on your score is not good and I should be getting more, but I just ran BB once and continued on setting everything else up.

It wasn't until I went back to try running it again where It would crash almost all the time. I made some recordings of the BB running with different settings (P core ratios, XMP off, disabled E-cores blah blah)

Thing is the system just wasn't stable (but only in BB) I could run cinebench forever without issues.

When I get home I"ll do a detailed analysis but I think the issue is related to the process scheduler (Thread director) in Windows 11.

On times when the system didn't completely crash (and i'd lose my recording) The system would stop loading the P-Cores and only use the E-Cores. Other times, despite having the max number of threads requests, not all the P-Cores would be loaded on my CPU.

I then ran and recorded BB on my i7 12700K (8P 4E) CPU and had zero issues. I also ran it on my laptop i9 13900H (6P 8E) without any issues.

Its a topic for another thread, but in a similar type of performance regression. I can spawn 40 cars and drive around and get 40-60fps......then all of a sudden the FPS drops to 10-18fps for a while, then jumps back up to 40-60fps.

This behavior is mirrored in the BB on certain runs.

I can't wait to get home because I have a pile of data to go through to figure this problem out.

 

That is interesting. Even more interesting as it seems happening only on 14th gen?

 

Since you have a Z motherboard, can you try beamng with hyperthreading disabled. With my 14700k, the framerate stayed similar but the lows went up. I'll thoroughly re-test it after the exams are done(I am a student).

 

Ryzen 5800X3D - Vermeer - 4.427 GHz - 8c 16t
- Max Mbeam/s: 492.255
- Mbeam/s/Core: 61 523
- Mbeam/s/Core/GHz: 13 899
- Max cars before <100% Realtime: N/A (benchmark stopped at 40@120%)

Ryzen 5600X3D 6c 12t (Simulated)
- Max Mbeam/s: 392.19
- Mbeam/s/Core: 65 365

Ryzen 5600 4.37GHz - 6c 12t
- Max Mbeam/s 366.817
- Mbeam/s/Core: 61 136

 

Well that pretty much confirms that 3d cache doesn't do anything for Beam.......

 

This is going to be a really long post after a few days of staring at BB 100's of times trying to figure out what's going on with my computer as I don't think i'm getting the scores that I should (even though its just a score) as the game plays perfect for how I want.

Specs
MSI Z790 Pro Wifi
i9 14900K
Corsair H100 Elite AIO Cooler
32GB Corsair Vengeance DDR5 6400
Asus ROG Strix RTX 4090
1TB WD SN850x HS

When I first got the system together, CPU wise I was running 5.8ghz all core (P Cores) and 4.4ghz E Cores)
I basically let the motherboard go hog wild with power limits and voltage

This was my score


I didn't really look in to it much but it seemed alright. All my other benchmarks seemed in line (Time Spy, Cinebench, Passmark)

This was my highest CB23 Score



With the above settings I was getting some thermal throttling but only during things like Cinebench during gaming and BB no throttling.

Once I found this thread and found some others (14700K and a 13900K) getting slightly higher scores I had to wonder why I was not getting slightly higher scores.

I then started tweaking settings and that's where BB started to get real flaky.

When running BB 1 of 4 scenarios would happen.

1. The benchmarks would run and it would be a "clean" run, meaning it would load the CPU cores up and the results would be consistent (meaning no large drops in M/Beams between close numbers (aka small deviation going from 33 to 34 M/Beams

2. The benchmark would run and it would be a "dirty" run with scores much lower and deviating wildly between similar M/Beam numbers

3. The benchmarks would crash and the computer would crash

4. The benchmark would only get to a certain number of M/Beams and then "finish" without giving a score.

Referring to scenario 1. Here is a video showing a relatively clean run and I also included CB R23 as well to show that there is no thermal throttling going on.

Settings for this video are 5.7ghz P Cores 4.6ghz E Cores 400W PL1 and PL2 limits



Really no issues, the score is a bit lower than I would like but I"m also running HWinfo with a high sampling rate to try and show how the cores are being loaded as well as CPU temps

Without posting a ton of useless videos i'll post screens of the BB now for a "dirty run"





The first run scored lower than the second, motherboard settings were 58X multiplier 5.8ghz P Core and 4.4ghz E Cores. Some how with an adjusted multiplier the benchmark completed.

The second score wasn't bad however the consistency was not there as the M/Beams increased like the first BB result I posted. I've posted that picture below



Neither one of the benchmarks caused my CPU to thermal throttle, temperatures just aren't an issue, so i'm not sure why i'm getting such a variation in scores from run to run.

Scenario 3, well I can't post videos of that because they didn't turn out as the computer crashed while recording.

Scenario 4 looks like this



In this case it ran fine and then the P Cores dropped out and the E cores only got loaded.....then the benchmark ended and didn't give me a score.

I'm not really sure if its a bug with the 14th Gen as there is another person with a 14700K that doesn't have any issues. I'm leaning towards something with motherboard settings, but i'm open to suggestions.

I also played around with E Cores and P Cores being disabled.

With all the E-Cores disabled....it wouldn't finish



1P Core and 16 E cores



Hyperthreading disabled



Now this is where it gets interesting, the HT disabled was not a super low score, in fact it was in the range that I was getting sometimes during "dirty" runs with HT enabled.



Virtually the same score, HT vs no Hyperthreading.

One more to finish off, here is an 80 car BB test.



I'll start another thread about the odd performance regression when spawning a bunch of cars (40+)

 

MB: MSI Z790 Pro Wifi
CPU: 14900k
Cooler: Artic liquid freezer 2 (360)
Ram: 64gb @6000 CL 30

Completely stock cpu settings apart from Intel Enhance turbo turned off

 

I think @Tyler-98-W68 is onto something with the performance regressing unusually on intel processors and i think it has something to do with win11 thread director.

To test my hypothesis I installed win 10 enterprise on a separate sata ssd. After the install i booted into win10 and copied beamng drive to said sata ssd from my main win 11 nvme ssd.
upon running banana bench on win10 you can see that the single car score is obviously much lower since win10 or banana bench doesn't know how to schedule the thread to p-cores and instead 1 car runs on a e-core causing the 1 car score to be 63 MBeams/s. (IMG 1)
But the overall score is much higher in win10! as you can see i hit 1083MBeams/s in win10 with every car being added to the benchmark causing said mbeam score to increase linearly until 27 cars, while in win11 the mbeam score starts higher as it should since bananabench is only using 1 p-core resulting in a 1 car score of 106 Mbeam/s but performance weirdly regresses after 7 cars and regains unpredictably until 28 cars resulting in a score of 983 Mbeam/s in win 11. (See previous posts of mine all using win11)

I have experimented more than I wrote here trying to use process lasso to bind affinity / cpu sets or using unpark but none of those work. I tried this because I see some p-core threads being parked when bananabench is running in win 11 which obviously shouldn't happen.

So I don't know what now? Should I start a separate thread as a bug-report? Does anyone else know of a better solution? Share your experiences as well because this is very odd and using win 10 I blew my previous score out of the water lol.

 

The interesting part about your results (and your theory is spot on)
Is that on a "good run" my 14900K is showing higher consistent scores in the higher car counts (say above 32) vs your 14700K which makes sense because I have more cores.......Ideally I should be getting the highest scores in the 28-32 thread range (as I have 32 threads)

So i've attached this, back to back runs.....and i've highlighted in Green the "good" parts of the test, and red as the bad parts of the test, its funny because the "higher" score test had worse consistency than the lower score test.

 

Here is my Stealth 15m
i9 13900H 6 P Cores 8 E cores

Notice the performance consistency, everything is almost the same, a variation of only about 100 M/Beams, much different than my i9 14900k

 

NEVER MIND THIS, I think E-cores are supposed to have a higher MPP so tasks primarily go to E-cores first in background.



The plot thickens! It could be something firmware related to 14th gen processors!? I will explain.

So I looked online about thread scheduling to cores and preferred cores etc. I found that you could see what the preferred cores are according to windows. And it seems like windows is preferring E-Cores!

How do I know?

1 - Reboot PC (It's easier to find the event then)
2 - Open event viewer
3 - Click on Windows Logs then System
4 - look for a bunch of events equal to the amount of threads you have with Even ID 55
5 - Hold ctrl select all the events and copy as text
6 - Paste into notepad and save the txt file.

As you can see in my case the E-Cores have a higher reported Maximum Performance Percentage (MPP) than the P-Cores in win 11.
But the problem is that is also true in windows 10. So I don't know if this actually matters in the end but that's why I hope others here can share their reported MPP for their processors, even if you don't have E-Cores.
Supposedly you can't change the reported MPP scores because they are reported to windows by the cpu microcode so we can't see if changing MPP actually affects bananabench or not.

(I also recorded how you can quickly copy your reported MPP score incase my written tutorial is worded weirdly)
--- Post updated ---

Were you using win10 or win11? Can you also report your MPP score? Because my score in win10 is only 10mbeam/s lower than yours (post 1468) and I don't think 4 ecores are only supposed to be worth 10mbeams so you should be getting like 200mbeams higher than my 1080mbeams/s score

 

I doubt this benchmark and the real game are identical.

Anyways:
7800x3D Air cooled, 74C peak in this bench
DDR5 6000mhz CL30

 

Actually it does, the higher the M/Beam score, the more vehicles (AI) you can run at once without a slow down. Beam likes cores, it doesn't care much about 3D Cache, even though everyone thinks it does.

a 5800x3d got 492M Beams.......your 7800x3d got 619 m/beams. Not much difference, and likely attributed to memory speed increases (beam loves memory speed)

My 12700K at 5.2ghz P core speed gets 595

My 14900K gets 985m/beams.

When I spawn a ton of cares with my 12700K it slows down........when I use my 14900K......it slows down, but not as much, so yes your 3dCache isn't doing anything for the game.

Example.

You spawn 40 AI cars and tell me how your game runs

12700K



14900K



30 fps vs 50 fps, a very tangible difference in performance

 

This test is complete nonsense!) I compared the performance in the game at minimum settings, without emphasis on the video card! 13900k and 13600k, the difference in frames was minimal, although in the test it income up to 30% as well as the test shows a higher score with hypertrading enabled! In the game itself it only reduces performance!

 

And I see the opposite i've had progressively more powerful computers, and the higher the M/Beam score, the better the performance in the came with large amounts of car/AI spawned

 

idk if I did it right. is this good or very bad?

 

It's only somewhat accurate if you spawn 40 AI cars.
As far as i can see this benchmark only does specific physics calculations, not including reflections and all the other ingame things CPU does.
13600k has 20 threads, 13900k - 32, so with less than 20 cars the real difference will be minimal but benchmark tests 40.
--- Post updated ---

I play almost exclusively Beammp where most servers have a limit of 10 players, at 10 cars in banana i got 450 which seems to be higher than 14900k despite the frequency difference so the cache definitely does... something.
--- Post updated ---
Higher than some 14900k*

 

As I said here Bananabench does not directly represent the in game performance.

 

If you want to see if 3D Cache does anything for gaming, Download and Run City Of Los Injurus, its CPU dependent and doesn't really touch the GPU much, I've done multiple systems and logged each run with Los Injurus, but again I still see no compelling evidence 3Dcache does anything for Beam

 

I was using windows 11. I'll attach my event log (I hope I did it right)