Performance with Intel Xeon

So, since the game is more CPU Bound, I would like to know if there are people with any Xeon platform (and which model) and how is the performance of the processor in this game in relation to the use of the Cores/FPS and the percentage of use. To find out if the game takes advantage of the various Cores/Threads present in these processors or if the IPC is even more important

 

So having benchmarked BeamNG across multiple CPUs,

• Platinum 8176 x2 56c/112 (Asus Sage C621E)
• Platinum 8180 x2 56c/112 (Asus Sage C621E)
• Platinium 8280 x2, 56c/112 (Asus Sage C621E)
• 2696 V4s, 48c/96ht (Some Chinese X99 Dual Socket)
• 10980XE, (Intel HEDT, forgot which 2066 platform specifically.)

All of these were ran with 256GB of DDR4 @ 2066, minus the X99 was 128GB at 1800 and default RDIMM timings. Forgot which latencies, sorry.

The dual socket platforms were ran with HT off, and NUMA off between tests. No differences in any BeamNG performance.

And consumer desktops for reference,

• 12700K
• 13900K
• R9 3900X
• R9 5950X
• R9 7950X3D

That these Xeon Scalable(Skylake) platforms do NOT at all offer the single core performance in comparison to even older Skylakes, the older Sandybridge Xeons performed much better than Skylake series, probably because of the MUCH MUCH larger L3 cache and boost clocks. Talking even as far back as 6th gen.

The mbeam benchmark is not conclusive, and is NOT at all a good reference target for Xeon or older intel HEDTs. You can do 40 cars at 480M/beam, but you're going to get 20 FPS. Sandybridge 2696 performed on par with the 3900X, but large margin of error running server CPUs and boards.

The 2696 was only outperformed by the 8280 due to much higher boost clocks. At the caveat of pulling 350-400 watts, (200ish per socket.) and various settings adjusted in the C621E's BIOS to make sure it was constantly pushing boost clocks, and pushing the two chips to constantly reach thermal throttle at 70c on water (The two 8280s). I only managed to test these because of my own actual curiosity as well, and sourced them from vendors looking to upgrade to EPYC systems.

I don't even want to get into the argument of two socket versus one socket on server boards. NUMA and other windows threading assignments, even disabling a socket and running one singular Xeon chip is not going to get any performance gains. There just isn't the IPC/clocks/latency/timings/chipset/I/O inside the other Xeon chips in comparison to Ryzen and Intel's 12th gen.

I can't comment on the newer W9s.
Sandybridge Xeons were great, and still are for budget platforms.
Skylake Scalables, Skylake in general was pretty poor for performance.

I can go ahead and run proper benchmarks again, but the only references I have now are the 8176, 2696, and 7950X3D. But having tried BeamNG on all of those which I've listed, the consumer desktops are ALWAYS going to perform better in real performance, even at 20+ vehicles in simple setting.

 

would like you to test the 7950X3D, with a say 4080 or near that.
either you can do in wcusa, or italy but I'd say wcusa is the best as theres no other limiting factors.

The tests I've been conducting have changed over multiple occasions since many of it dont meet the expectation or reality of the sorts. (my old benchmark would score higher than an actual traffic, geometry around the place is low/high, etc)
So anyway, would like to just see as many cars as the camera can see, physics on and frozen, and while all cars idle and moving at the same time. This is the kind of test that would give the best outcome for close to reality for benchmarking.

(Also would like to dm you for tips on my new benchmarking tool that is still wip)

 

Thanks for the excellent feedback. I was thinking of testing with an E5 2667v4 and an RTX3060ti 12gb and 32gb quad channel to see if the game would take advantage of 8/16 at 3.5ghz

 

Hey i would Love if you tested the 7950x3d with just spawning cars inside beamNG itself and telling me the performance or Banana Benchmark either would work but im looking at 13900K or 7950x3d and need the Truth on which to get for beamng

Oh and Can you tell me the performance of the 10980X with cars spawned? I'm also debating on 10980XE or 5950x (if i cant afford 13900K or 7950X3D for any reason)

 

I've taken a gander and purchased two Xeon 8480+ ES (SPR), and Xeon 8570 ES (EMR) two CPUs on a Gigabyte dual socket MS73-HB0 motherboard.

Largest take aways:

• 112/224 with all enabled tiles did not improve much for performance, BeamNG does not appear to scale across multiple sockets. With how EMR/SPR scale with high core count, disabling tiles and cores from 112 to 56 improved overall boost frequency from 2800 all core, to 3800+ all core.

• Disabling tiles for 28/56 per socket nets improved performance up to 40 vehicles, not having disabled tiles for 56/112 allowed scaling up to 99 vehicles.

This however only saw marginal improvement up to 40 vehicles, before degredation hit, a little bit faster than all 56 cores than 28 for up to 40 vehicles. (Because of boost frequencies of 3800 versus 2800.) This shows that BeamNG is indeed scaling beyond 28/56

• Both CPU sockets were not utilized, only socket 0 cores and hyperthreading up to 91 "cores" being utilized by BeamNG.

Is it worth buying server grade Engineering Samples/$2,000-10,000+ CPU/Multiple socket for more vehicle density? No, the game's engine appears to have deminishing returns with high core counts, and does not scale with multiple sockets.

Why Xeon over EPYC?

• Xeons in the price range for these engineering samples offered much higher boost frequencies than EPYC processes did of similar price. The game very much favors higher clock frequency with total core count, or total cache sizes. SPR/EMR chips have excellent/similar IPC versus 2nd or 2rd Gen EPYC.

Would a Threadripper 7000 be better? Or Intel Xeon W W9-3495x/3595x?

• Definitely. Very much definitely. High core clocks with high core count.

 

This was an excellent test, my doubt before was to have some performance improvement using Xeons V4 from the LGA2011-3 era. But apparently the game requires more IPC and Clock than the number of Cores

 

I just put together an old Dell prescision 5810 tower for my Nephew to play on, swapped the Xeon E5-1620 v3 (4c/8t) for a E5-2698 v3 (16c32t) and 64gb DDR4. It runs it surprisingly well. I have no data to go off of but it loads the menus and traffic just as fast if not faster than my 10400 with 32 gb DDR4. V3 and V4 xeons are CHEAP and that influenced the build quite a bit. I spent maybe $120 on the tower, $30 on the CPU, and threw in a 6650xt that I had laying around.

 

Yeah, I was really keen to put together a kit with X99 + Xeon 2640v4 or 2667v4 with 32gb in QuadChannel and some RTX XX60 series

 

Not sure what's happenend, but since 0.34 BeamNG has not been happy with Vulkan API being used with Platinum 8480+ or W9-3495X's. CPUs will not turbo to 3700 or 4800 respectively. The game is still playable in DX11.

 

I've went ahead and brought down my other workstation with a W9-3495X and the performance difference is significant, even though they're the same architecture/SKUs, just the W9-3495X is an unlocked 8480.

For testing simplicity and not skewing results, minus the dual socket config, both single socket 8480+ and W9 are on a BIOS edited ASUS WS W790E to allow the 8480+ to boot on it.

Single socket 8480+ (No cores disabled 56/112).
All core turbo max with no disabled cores is 2800
Disabled to 28/56 max turbo is 3800
Max turbo freq. across 28 cores at any time is 3800.

• ASUS WS W790E
  
• 8x 32GB 4800 DDR5 LRDIMMs
• 7900XT
• 990 Pro

W9-3495X (No cores disabled 56/112)
All core turbo max with no disabled cores is 4600
Disabled to 28/56 max turbo is 4800
Max turbo freq. across 28 cores at any time is 4800.

• ASUS WS W790E
  
• 8x 32GB 4800 DDR5 LRDIMMs
• 7900XT
• 990 Pro

Dual socket 8480+ (No cores disabled 112/224) Filled out the DIMM slots on this to achieve full channnel.
All core turbo max with no disabled cores is 2800
Disabled to 56/112 max turbo is 3800
Max turbo freq. across 28/28 cores at any time is 3800.

• Gigabyte MS73-HB0
• 16x 32GB 4800 DDR5 LRDIMMs
• 7900XT
• 990 Pro

Testing is with 46 vehicles/max pool, or no vehicles. 46 vehicles is to allow 10 cores at any time to rotate between parking. Virtual NUMA is enabled on all three, NUMA is enabled on dual socket config, but was also disabled for 1 test interval. This test was on Utah with D-Series for player spawned, and never moving and default camera upon spawn.

Single socket 8480:
46 vehicles/max pool
Average 20.1 FPS. Had lows into 10 FPS.

Single socket W9
46 vehicles/max pool
Average 41.7 FPS with occasional lows to 20 FPS.

Dual socket 8480:
46 vehicles/max pool
Average 14.1 FPS with lows of 11 FPS.

Further tested disabling cores between the 8480s to achieve max turbo freq of 3800 between both single and dual socket. The dual socket system still had hardware NUMA enabled, as well as cache prefetching from opposing core thru Intel's UPI link tech... There was no improvement/degredation between either of them save lower minimums with hyperthreading off load, the same applied to the W9 with cores disabled to 28 as well. With the next test I went ahead and disabled NUMA, left vNUMA enbled and further disabled cache prefetching between both L1/L2/L3 caches, and the game performed worse than the first test with it. This changed with using process lasso... With process lasso managing which cores BeamNG was utilizing, it gathered same performance as the single socket system, as nothing was offloading into the other socket and lack of NUMA. However, BeamNG's UI system wasn't changing. When the UI is opened it opens another PID (randomized) that I'm not able to forcefully assign, and the game tanks and WIndow's thread manager becomes a lost cause. But this isn't anything particularly wrong with BeamNG. There was some more BIOS adjustments done to test this, and after about 8 hours of constant restarts, pretty much met all what could be done.

Vulkan API was a disaster. The game is entirely unplayable (1-5 FPS) with dual socket configurations, and I am not able to figure out why. This happens whether NUMA/vNUMA, hyperthreading, cache prefetching- everything. I met this same issue on my 8570 EMR build too. To further isolate the issue if it's just dual socket bound, or these EMR/SPR systems I tried older dual socket 8176 Platinum system, alongside 2696V4s and it ran into the same issue with the Vulkan API.

Performs just fine on the W9-3495X, or single socket server configs.

Wish I had some Threadrippers to test. Game loves higher clock frequencies, regardless of decade old architecture, really.