BANANA Benchmark Results Comparison

with my upgraded system, i am getting about 160 mbeam/s with a stock i5-6600K...will try overclocking and see whats up after that

 

4.6Ghz on all cores and 4 cores active, 8th gen i7 gives this:


There is hardly any difference with how many cars I can run at 60fps if I have 6 or 4 cores active, depends a lot from settings and map of course.

If one has 6700K for example, upgrade to newer ones is not giving much, added cores don't seem to help much, especially if it means lower clocks it seems to be only nice numbers added, but amount of cars while maintaining 60fps is not changing really.

C1 highway map, I tested with 6 cars and with 4 active cores @4.6Ghz like in above banana bench, I got 55fps while 6 cores @ 4.6Ghz gave 59fps, GPU was not even at 60% load, so anyone wanting to run more vehicles and having 6700K overclocked, be warned that there might not be quite as much gains that some might lead you to believe.

If you are okay below 60fps, then more cores allow running more cars with less fps drop, so there is that, but to utilize more cores your one core need to be enough strong as each car adds load to that one core too, which then makes everything bit more complex.

On gridmap I can run 10 D15 pickups at 60fps with 6 cores active, but that is the thing, they are all the same vehicles, put in different ones and numbers drop faster, so be aware of these aspects which might be surprising.

 

That's a pretty healthy 74% overall improvement over my 2nd gen I5 (@4.2ghz). And a 40% single threaded improvement too. Not that I would expect less given 7yo I5 vs modern I7, even with the modern CPU held back to same number of cores active.

At a guess I would imagine the difference between the multicore improvement(74%↑) and single thread improvement (40%↑) is predominantly the hyperthreading at work (since 4 threads vs 8 threads). Which is interesting to see.

-- BeamNG technology speculation below --
I am now rather intrigued to know how BeamNG handles its multithreading. Since the devs have said that parallelisation within vehicles isn't workable. From a high level viewpoint, BeamNG's workload looks like it would be quite suitable for it. That said, with all nodes and beams being interdependent on each other, I can imagine that causing major headaches to get working, potentially without performance gains due to that interdependencies.

I would imagine the workflow of simulating a vehicle looks something like:

1. Calculate current trajectory of nodes
2. Calculate the stresses that movement would apply on the beams
3. Calculate the movement of nodes, taking into account the forces generated by the beams in step 2

If that were the case, and each step was calculated for the entire vehicle discretely (with a thread synchronisation after each one), it would make sense that steps 1&2 could be divided between as many cores as needed, with all cores working on a shared dataset. But I'm not sure how I would approach that 3rd step. I would guess that a Boids style approach where all the forces from the beams are combined into a new position would do it. If so, that could also be divided between multiple cores.

This leaves me wondering what I have neglected to consider, since my understanding suggests that it may well be possible and beneficial to multithread within vehicles. But ofc, my understanding of BeamNG as a piece of software (and the design choices taken in creating it) is essentially nill compared to the people who actually programmed it
My absolute best guess would be that its just not worth it for the devs considering that all the official content runs on a single thread just fine, even on middling 7 year old hardware. Hence why bother?


Of course, the above may not be based in reality, because I don't know how BeamNG functions internally. Obviously there is a lot more to it than what I wrote above.

 

Graphics, LUA, sounds, UI etc. "stuff" that is where main limit is (those or at least some of those are on single same core, each taking a piece), but of course with more complex vehicles we are seeing limit of what one core can do in physics, but that is not great issue currently, except with some mod vehicles like articulated bus.

There are of course many other considerations, for example keeping physics running 2000Hz realtime, now imagine there is some piece of LUA coming to say my turn you wait, that won't do so it makes absolutely sense to keep those other than physics away from cores where physics run.

What is odd, is that you can have trailer hooked up and that is using different core to my understanding as it is another vehicle. So what would prevent using couplers for parts of vehicle and sharing load that way? Hack of course, but isn't hacks what we live for

Anyway, considering official content and load physics cause, it is well possible to run two heaviest vehicles on single core in terms of physics. However physics is not the limitation, it is that other stuff, add vehicles and observe drawcall count, compare adding many of same and many of different, there are many challenges and it is not quite as simple as one may think, so developers definitely know their stuff and have very correct path in this, but it is only fair for users to note that because of way game works, it is wise to look hardware that has also good per core performance as many cores will not do much better if one is maxed out with 4 cores loaded.

Oh and remember that I was running at 4.6Ghz which is massive 0.3Ghz over what CPU runs at stock with all cores enabled most of the time, unlike for example 6th generation which spends much more time at higher boosts even when there is added CPU load. (some people refuse to accept that operational change existing, but they just don't know their stuff).

In practice setting all cores to slightly higher clocks gives nice boost.

You can open performance tool in game and check which eats most of your CPU time, usually it is Render and pre Render, that is graphics stuff, turn shadows off and see massive drop there, it is nice tool and helps a lot in evaluating which upgrade path to take, but takes bit of experimenting to get most of the tool.


6th gen non K 6700 gave me 40.552 for single car and 166.370 total, that is 4Ghz max turbo and 3.7Ghz all cores fully loaded, so 8th gen is not huge improvement, but that extra speed in single core does a lot. Something like 55-60Mbeams/s for single car would probably be enough to have more use of those extra two cores, but there really is nothing yet that can do that, but there are lot of CPUs with more cores, will be interesting to see what improvements BeamNG devs manage to find out, probably after year or two it is again different situation and potential of modern CPUs can be utilized better.

I would wait that 1-2 years until making upgrade for next 5+ years, I might be wrong in next estimate, but kinda would think that fast single core requirement will stay, you see CPUs are still improving that area too and developers can count it to take care of some of the required improvement, some optimizations will certainly also come, but it is likely that top of the line CPU models will always have edge on this game, but on that I can be wrong, I don't know future, I don't know plans of devs etc.

My next upgrade will be what CPU will give 30% or more single core upgrade over current overclocked performance, might be long wait

 

With rumors of Zen2 being able to clock up to 5.0Ghz stock for top models we might soon see an 8c/16t processor with an IPC above intels 8th gen, and with higher clocks too. I wonder how beam will be optimized for lots of high-speed cores.

Obviously not that 30% single core you want, but I don’t think there will be a major jump in clock speeds for a while after this 5.0Ghz barrier is broken.

 

That may well be the point where I upgrade from SandyBridge. Since I will be upgrading to Windows 10 on my main PC some time in the next year (fresh install, not sure I would be comfortable upgrading a 7 y/o Win7 install), it makes sense to get the CPU/MB/RAM out of the way at the same time. Perhaps with some NVME goodness to boot

Faster code compilation times would be pretty sweet. Will be waiting to see what both AMD and Intel bring to the table, and what makes the most sense for me when push comes to shove.

 

I'm actually more interested to see what happens in software, Zen2 if rumours are true certainly will be nice, but software needs to change for that to be practical improvement as with current software it would be still extra cores doing nothing, hardware is not going to offer jumps indeed and even current hardware can do so much more if it is possible to change software to use potential computing power better.

Zen2 will be great, there is no doubt about that, but I look further than that in hardware as software changes are probably going to give lot of boost.

You guys remember how in Playstation they could utilize only fraction of power at beginning, then as years did go by, they could harness much more of computing power, kinda similar situation in software world currently.

It has been quite nice change so far, but there is still so long way to go.

 

well, I just bought a used server, so I thought I'd try benchmarking it against my desktop. I've probably posted my desktop with the same hardware here before but it's definitely years out of date.

2x Xeon E5-2660 locked at 2.7GHz, HT On

Code:

][ v6 | 0.5.2.0 (64 Bit) built Sun Feb 17 13:52:38 2019 | now: Saturday Sat Apr 13 02:39:12 2019
][ Windows 6.2 (build 9200), 64-bit | RAM: 55.97 GB
][ Intel(R) Xeon(R) CPU E5-2660 0 @ 2.20GHz | 32 cores | 2200 MHz
 *** pickup ***
 # | Dynamic Collision ON| Dynamic Collision OFF
   +----------+----------+----------+----------+
   | MBeams/s | % Realt  | MBeams/s | % Realt  |
---+----------+----------+----------+----------+
 1 |   21.468 |   267.75 |   27.967 |   348.81 |
 2 |   41.385 |   258.07 |   54.294 |   338.58 |
 3 |   59.950 |   249.23 |   80.009 |   332.62 |
 4 |   79.505 |   247.90 |  102.804 |   320.54 |
 5 |   97.825 |   244.01 |  133.250 |   332.38 |
 6 |  114.012 |   236.99 |  154.392 |   320.93 |
 7 |  134.051 |   238.84 |  181.394 |   323.19 |
 8 |  140.482 |   219.01 |  172.423 |   268.81 |
 9 |  110.468 |   153.08 |  144.883 |   200.77 |
10 |  117.888 |   147.03 |  155.005 |   193.32 |
11 |  124.326 |   140.96 |  167.469 |   189.88 |
12 |  133.180 |   138.42 |  178.600 |   185.62 |
13 |  141.066 |   135.34 |  190.634 |   182.89 |
14 |  146.814 |   130.79 |  200.179 |   178.33 |
15 |  155.987 |   129.70 |  211.854 |   176.15 |
16 |  164.018 |   127.85 |  216.500 |   168.76 |
17 |  165.014 |   121.06 |  221.022 |   162.15 |
18 |  173.791 |   120.42 |  233.108 |   161.52 |
19 |  173.590 |   113.95 |  241.715 |   158.67 |
20 |  182.346 |   113.71 |  249.876 |   155.82 |
21 |  190.318 |   113.03 |  252.180 |   149.77 |
22 |  193.412 |   109.65 |  268.885 |   152.43 |
23 |  203.440 |   110.32 |  265.582 |   144.01 |
24 |  191.564 |    99.55 |  245.143 |   127.39 |
25 |  187.592 |    93.59 |  241.526 |   120.49 |
26 |  178.228 |    85.49 |  233.668 |   112.09 |
27 |  178.199 |    82.31 |  233.297 |   107.77 |
28 |  174.237 |    77.61 |  232.387 |   103.51 |
29 |  179.404 |    77.16 |  227.896 |    98.01 |
30 |  177.243 |    73.69 |  224.589 |    93.37 |
31 |  171.900 |    69.16 |  225.256 |    90.63 |
32 |  173.781 |    67.73 |  221.870 |    86.47 |
Max Mbeams/s:   203.440 Mbeams/s

2x Xeon E5-2660 locked at 2.7GHz, HT Off

Code:

][ v6 | 0.5.2.0 (64 Bit) built Sun Feb 17 13:52:38 2019 | now: Saturday Sat Apr 13 02:48:37 2019
][ Windows 6.2 (build 9200), 64-bit | RAM: 55.97 GB
][ Intel(R) Xeon(R) CPU E5-2660 0 @ 2.20GHz | 16 cores | 2200 MHz
 *** pickup ***
 # | Dynamic Collision ON| Dynamic Collision OFF
   +----------+----------+----------+----------+
   | MBeams/s | % Realt  | MBeams/s | % Realt  |
---+----------+----------+----------+----------+
 1 |   21.963 |   273.92 |   28.990 |   361.56 |
 2 |   42.872 |   267.35 |   55.633 |   346.92 |
 3 |   61.258 |   254.67 |   80.692 |   335.46 |
 4 |   80.300 |   250.37 |  105.473 |   328.86 |
 5 |   99.045 |   247.06 |  136.049 |   339.36 |
 6 |  118.149 |   245.59 |  162.068 |   336.88 |
 7 |  121.433 |   216.36 |  157.735 |   281.04 |
 8 |  130.198 |   202.98 |  164.993 |   257.22 |
 9 |  138.149 |   191.44 |  180.249 |   249.78 |
10 |  147.980 |   184.56 |  193.530 |   241.37 |
11 |  155.557 |   176.37 |  204.957 |   232.38 |
12 |  167.219 |   173.80 |  217.664 |   226.22 |
13 |  174.875 |   167.77 |  230.473 |   221.11 |
14 |  182.069 |   162.20 |  237.366 |   211.46 |
15 |  188.402 |   156.65 |  248.624 |   206.72 |
16 |  186.912 |   145.70 |  240.329 |   187.34 |
17 |  151.597 |   111.22 |  201.350 |   147.72 |
18 |  156.328 |   108.32 |  211.458 |   146.52 |
19 |  160.726 |   105.50 |  218.094 |   143.16 |
20 |  165.164 |   103.00 |  227.905 |   142.12 |
21 |  171.857 |   102.07 |  232.054 |   137.82 |
22 |  175.511 |    99.50 |  237.817 |   134.82 |
23 |  177.439 |    96.22 |  240.743 |   130.54 |
24 |  177.610 |    92.30 |  242.818 |   126.18 |
25 |  181.883 |    90.74 |  244.506 |   121.98 |
26 |  184.983 |    88.73 |  247.908 |   118.92 |
27 |  184.868 |    85.39 |  248.314 |   114.70 |
28 |  189.918 |    84.59 |  255.122 |   113.64 |
29 |  194.126 |    83.49 |  259.718 |   111.70 |
30 |  194.471 |    80.85 |  257.552 |   107.07 |
31 |  194.253 |    78.15 |  262.439 |   105.58 |
32 |  197.187 |    76.85 |  265.591 |   103.51 |
Max Mbeams/s:   197.187 Mbeams/s

1x i5-4690k, locked at 4.6GHz

Code:

][ v6 | 0.5.2.0 (64 Bit) built Sun Feb 17 13:52:38 2019 | now: Saturday Sat Apr 13 03:11:46 2019
][ Windows 6.2 (build 9200), 64-bit | RAM: 15.95 GB
][ Intel(R) Core(TM) i5-4690K CPU @ 3.50GHz | 4 cores | 3500 MHz
 *** pickup ***
 # | Dynamic Collision ON| Dynamic Collision OFF
   +----------+----------+----------+----------+
   | MBeams/s | % Realt  | MBeams/s | % Realt  |
---+----------+----------+----------+----------+
 1 |   45.695 |   569.91 |   65.693 |   819.31 |
 2 |   89.879 |   560.48 |  125.063 |   779.89 |
 3 |  124.124 |   516.02 |  179.736 |   747.22 |
 4 |  153.758 |   479.41 |  209.390 |   652.87 |
 5 |  116.708 |   291.12 |  164.397 |   410.07 |
 6 |  131.040 |   272.39 |  183.310 |   381.04 |
 7 |  148.295 |   264.22 |  203.493 |   362.56 |
 8 |  149.010 |   232.31 |  189.913 |   296.07 |
 9 |  129.137 |   178.95 |  177.103 |   245.42 |
10 |  135.572 |   169.09 |  185.569 |   231.44 |
11 |  137.755 |   156.19 |  188.077 |   213.24 |
12 |  133.653 |   138.91 |  179.956 |   187.03 |
13 |  124.405 |   119.35 |  170.327 |   163.41 |
14 |  123.452 |   109.98 |  168.378 |   150.00 |
15 |  122.773 |   102.08 |  166.646 |   138.56 |
16 |  120.594 |    94.00 |  151.519 |   118.11 |
17 |  113.618 |    83.35 |  148.288 |   108.79 |
18 |  112.814 |    78.17 |  150.220 |   104.09 |
19 |  110.817 |    72.74 |  145.119 |    95.26 |
20 |  107.598 |    67.10 |  136.141 |    84.90 |
Max Mbeams/s:   153.758 Mbeams/s

I observed some... odd? CPU usage on the server, where it'd fill up one CPU in entirety, then the other, regardless of hyperthreading being on or off. Blame BeamNG? Blame Windows' trashy scheduler and handling of NUMA nodes?

For reference, in Cinebench, the server with HT on scores ~1800, HT off scores ~1400, my desktop scores ~680

It's also likely worth mentioning I have spectre/meltdown mitigation disabled on both machines.

Clearly if scaling was in line with other programs, we should be seeing at least ~350MBeam/s on the server side. However, the CPU usage was quite jumpy and never reached 100%, even when running 16 vehicle threads on one CPU. Bananabench will reach 100% usage on my desktop. I haven't actually tried RUNNING beamng on the server, because duh, we all know what happens with that many cars in-game. No point.

Also interesting to note, of all the things M$ locks to Windows Server or Windows 10 Pro for Workstations, dual socket motherboard support is not one of them. ¯\_( ͠° ͟ʖ ͠° )_/¯

 

i7 4770k with 3.9GHz Turbo. I expected less tbh, but it definetly is an upgrade to the 50 or 60 score my old FX4130 got.

 

Here with a small Xiaomi Air 12 with a m3-7y30

 

Here is something interesting, 115.364MBeams/s with i7-8086k that is set to 3Ghz 4 active cores and no HT so that is 4 Threads only. Those who have old i5 CPU around 3Ghz can compare and see how much or little IPC gains there has been during the years for Intel CPUs:

You can compare results for example to post of @aljowen few posts above, pretty close same kind of scores, but his i5 has 1.2Ghz higher clocks I believe?

Another point of interest, during Banana bench I saw 26.782W maximum power usage by CPU, highest temp of any cores during banana bench was recorded as 35C

Also Z370 Asus MB has higher VRM temps at idle than this CPU can reach on any stress with this kind of settings, but of course VRM temp is not exactly getting high either with this low power consumption.

 

I just ran the tests again, I am topping out at 61c across all 4 cores, I would guess they overall average around 59-60ish deg c. That is with a 12cm air cooler, without the PC getting any noticeably louder. The thermal paste is 8 years old, but I don't have any reason to suspect that it is causing any issues there. I have no idea what VRM temps I am getting, I could find out, but they are probably fine.

And yes, 1.2ghz more than 3ghz. My CPU sits at 4.2ghz with no speed stepping, since that causes blue screens for me.

So as a rough guess, that is probably around a 30-40% increase in IPC over 7-8 years? Certainly not spectacular, but with hyperthreading etc, certainly enough to make upgrading the CPU a very tempting prospect when Win7 gets closer to eol. Bring on Zen2 and whatever Intel counters it with (whether its new CPUs or price drops) xD

 

Also excellent 2nd hand deals are not to be forgotten, I suspect that within 6 months there will be lot of offering and for example 6th gen intel stuff might go quite cheap while still delivering perfectly fine performance figures for BeamNG.

Zen2 should be pretty similar clock to clock of what my CPU gets, but there is not knowing right now I guess.


I did also test 8086K set to same clocks as 6700 non K and got very slight increase, 177Mbeams/s over 166Mbeams/s which might fit into variation between runs. Considering that is difference of 2 generations there is not too much progress happened during that time which is several years of the that 7-8 years time frame.

This kinda underlines the need to move into more cores in hardware and software, if their ability to improve IPC has stagnated so bad, there is no other place to go really in order to keep more complex code possible into future.

Quite interesting to see these per clock speed changes too.

 

I wondering something, maybe it is windows 10 thing, but 1 car is 46 for you, I got 52.586 with 5Ghz, however that is highly sensitive to what is running on background.

Total with 5Ghz I got 301.477 but also on that same things affect that.

Quite small differences though, no practical meaning. Try sometime normal graphics preset without ssao or reflections and be amazed of few things

Update: 4.8Ghz is 1.15V so it is not much heat so I prefer that.

 

Another fun test is to run only with 4 cores, sure Bananabench result will be less, but there can be few surprises

 

There has to be some improvement in 0.16 as I'm now getting higher scores and all drivers and bios are same as before.

i7-8086K @ 4.8Ghz 6c / 12 threads 313.912Mbeams/s:


I have old (0.15 or 0.14) 4.8Ghz result too for comparison 290.699Mbeams/s:

 

Here is my 9920X @ 4.6 all cores. I had quite a few programs running in the background so I am sure I can get a higher score.