Naz Posted September 21, 2020 Share Posted September 21, 2020 First up some disclaimers and notes on the benchmarks. 1. These figures should be taken as ball park figures and not absolute values. 7DTD is a very difficult game to benchmark accurately. It's in alpha and sometimes does weird things. Also some things are difficult to account for such as the ai, they will behave differently and spawn differently every run, which leads to #2. 2. What's controlled for and what's not I haven't controlled for driver versions & windows versions. My goal was just to get "close enough" results instead of 100% accurate as possible. That said i have controlled for memory leaks, restarting the game after every run. Time of day is reset from the console after every run. When changing resolution i changed ingame before shutting down then validated with the unity screen selector on next launch. Background programs were closed for the tests. No zombies/animals where killed so areas weren't "cleared" until x amount of days, to insure that on following runs they would respawn (Although i can't do much if zombies kill animals or vise versa etc). I only did 1 pass on each test, however any result that didn't look right or didn't make sense was discarded and retested. 3. 2 systems isn't enough to draw definitive conclusions for every configuration, no AMD GPU's were tested, only windows for the OS ect. The conclusions found here may or may not apply to your own systems, these results are only really comparable if you have similar hardware. But they will however give an idea of where the current performance is at with the hardware that was tested. 4. It's Alpha Any update could change any conclusions drawn from these tests. Also as a work in progress things are always improving and getting worse, by the time 7DTD goes gold these results with be obsolete and invalid. Things are always changing in alpha. 5. Console Options used Since A17 I've always used sg optionsviewdistance 90 to increase the field of view to 90 degrees and sg optionsviewdistance to increase the view distance to 12. A18.somthing brought back the FOV slider but it's max is 85 degrees, i doubt the 5 degree less fov has any measurable impact on performance. So i stopped using it for a19, However all other tests are run with the increased view distance. 6. Tests on different resolutions were done on monitors native aspect ratio 16:9 Resolutions were tested on an Asus PB287Q 4K 60HZ Monitor 21:9 Resolution was tested on an LG 38GL950G 3840x1600 175Hz Monitor 7. The benchmark Run I've been using the same run since A14 for all tests, you can find more details on the exact run in my a15 benchmarks i did ages ago Here My Other 7DTD Benchmarks A19.3>A19.4>A19.5 Hotbox Benchmarks Resizable Bar Trident Benchmarks A19 RTX 3090 Trident Benchmarks A19.2 B3 Hotbox Benchmarks A19.0 Benchmarks 2 Systems Intel,AMD,SLI,Core Count And SMT Tested A15 Trident Benchmarks Systems Tested Spoiler "Trident" "Hotbox"CPU: AMD 3950x 16 Core 32 Threads CPU: Intel I7 6700K 4 Core 8 ThreadsMotherboard: Gigabyte X570 Aurus Xtreme Motherboard: ASUS STRIX Z270I Mini ITX MotherboardRam: G.SKILL Trident Z Neo 64GB (4 x 16GB) DDR4 3600Mhz CL16-19-19-39 Ram: Team Group Vulcan T-Force 16GB (2x8GB) DDR4 PC4-19200C14 2400MHz Storage: Samsung SM951 512Gb M.2 SSD Storage: 2x Kingston SA400S37/240G 250GB Sata SSD's RAID 1GPU: 2x Evga Sc2 Gaming GTX 1080ti Black Edition GPU: Palit Gamerock Premiere Edition GTX 1080 Systems Overclock Notes Spoiler "Hotbox" CPU: 4.5Ghz, 1.3V, Cache ratio 43 Ram: Stock: Default JEDEC(2133Mhz) , Overclock: XMP Enabled (2400Mhz) GPU: Core:+50 Mem:+325 "Trident" [All Core] CPU: Vcore: 1.3V CCD 0: CCX 0&1: 41.75 CCD 1: CCX 0&1: 40.75 Ram: Stock: Default JEDEC(2133Mhz) , Overclock: XMP Enabled (3600Mhz) GPU Single: Core:+50 Mem:+250 GPU SLI: Core:+20 Mem:+250 "Trident" [Gaming Profile] CPU: Vcore: 1.45V CCD 0: CCX 0: 45.50 CCD 0: CCX 1: 44.50 CCD1 DISABLED SMT: Disabled Down Core Control: Six + Six Ram: Stock: Default JEDEC(2133Mhz) , Overclock: XMP Enabled (3600Mhz) GPU Single: Core:+50 Mem:+250 GPU SLI: Core:+20 Mem:+250 2 Different video Settings Tested Spoiler Ultra Settings 1 Ultra Settings 2 Lower Settings 1 Lower Settings 2 Original Benchmark Excel Files Download Hotbox Benchmarks A19.0 B180 & A18.2 B5 Ultra Settings Spoiler A19.0 B180 & A18.2 B5 Lower Settings Spoiler Trident Benchmarks Trident specific note The system "Trident" has much more going on, so there is much more to test. We'll be looking at your standard tests ultra and lower settings like Hotbox, plus those tests with SLI. Also based on my tests 7DTD runs best with 4 true cores, no Hyper Threading or SMT . I didn't test that on Hotbox in it's current config, but in previous alphas before i put the I7 6700K in there i got much better performance with my I5 4670K than i did with the 6700K. So it seems to apply to both Intel and AMD CPU's. Also the 3950x/newer AMD CPU's are build a bit differently to Intel/Standard CPU designs. Normally there is just 1 die on the chip or a "monolithic die" all the cores, cache, io controller is all on one "chip". AMD's newer CPU's have multiple die's or "chiplets" the die's that have the CPU's core are called compute complex's "CCD's" and each "CCD" is split into multiple groups of cores (CCX's). They communicate with each other and other ccd's through a super fast bus called "Infinity Fabric". While these are very fast and have minimal impact on performance in some workloads. It's still slower than if all the processing was done on 1 chip. So with those 2 things in mind i made a "gaming profile" in the bios that i switch to for the best performance when gaming. One of the 2 CCD's is disabled and 1 of each of the remaining CCX's cores are also disabled, as well as Hyper Threading (SMT). This leaves 6 Cores and 6 threads all in one CCD. Although 4 cores is best my tests showed minimal loss in performance, so i chose 6 to make it more general, if i was playing another game that utilised more threads. With much more thermal and power available with this setting i was also able to get a higher overclock on this profile, All these things combined results in huge performance gains compared to having all cores enabled. Ultra Settings Single GPU Benchmarks Spoiler Lower Settings Single GPU Benchmarks Spoiler Ultra Settings 2X SLI Benchmarks Spoiler Lower Settings 2X SLI Benchmarks Spoiler Conclusions A18-A19 We can see that in higher resolutions performance has increased and lowered at lower ones. This shows us that there have been some GPU optimisations, this supports the claim of the new zombies being easier on your graphics card, so it should improve a little more when all types have been updated which is great. The losses at lower resolutions tell us that CPU optimisations have taken a hit this alpha, 7DTD is a massively CPU bound game and any losses in this area is going to have the biggest effect. This theory is confirmed by the substantial losses in the 1% and 0.1% lows across all tests, which also demonstrate the stuttering issues persisting in A19, as well as the losses at lower resolutions. Cores & Hyper Threading ON or OFF We can see massive gains going from the all core profile to the 6 thread profile. At 1080p we see anywhere from 20%-70% increase in performance. Although i didn't post the older results on Intel, i can say it helps but not to the same degree it does on tridents 3950x. The reasons for the massive bump in performance in this case were several factors. The higher OC we could get with less cores enabled. Forcing 7DTD to utilise fewer cores with 4 being the best in my tests. Disabling SMT no idea why 7DTD doesn't perform well with virtual threads, some other games are like that, but most games just don't show any meaningful difference on or off. 3rd Gen Ryzen cpu's different CPU Design performing better if we make 7DTD only use cores on 1 CCD. SLI There is no official support for sli in 7DTD so we need to rely on 3rd party profiles. That said for a voxel game the gains it can bring can be quite surprising. At 4K on the gaming profile we see scaling of over 49%. SLI in 7DTD always goes back to it's CPU Bound nature. It can scale positively and provide more performance but only if the CPU is strong enough. We can see this looking at scaling with the all core profile vs the gaming profile on 4k lower settings. SLI performs worse than a single GPU until we get to 4K and at stock it's a measly 2.2% gain vs the overclocks 13% gain. Meanwhile the gaming profile only scaled negatively at 1080p, 1440p overclocked and above all scaled positively with 4k oc scaling to 49% where we're now mostly GPU bound. CPU overclocking and fast ram are pretty much required for positive scaling. Link to comment Share on other sites More sharing options...
faatal Posted September 25, 2020 Share Posted September 25, 2020 We are definately CPU bound. Profiling last week after the terrain shader update and related tweaks, show that even more. Just tested a build and 4k res is running better than I've ever seen it at high quality, which is what I normally play at. I would be interested to see your numbers when 19.2 comes out. Thanks Link to comment Share on other sites More sharing options...
Naz Posted September 25, 2020 Author Share Posted September 25, 2020 3 hours ago, faatal said: We are definately CPU bound. Profiling last week after the terrain shader update and related tweaks, show that even more. Just tested a build and 4k res is running better than I've ever seen it at high quality, which is what I normally play at. I would be interested to see your numbers when 19.2 comes out. Thanks Thank you for jumping in. I've been following your post in the dev diary and it's great to hear about all these optimisations in the works. I'd be happy to do some follow up tests for a19.2, but it will just be "hotbox" and I've swapped out the 1080 for one of the 1080ti's i had in "trident" (Bought a 3090, flipped the 1080 and one of the 2 1080ti's then shuffled down the spare to the htpc ) I already have the stock numbers for a19.0 with the new card in and should have the oc numbers later today, so i should be ready when .2 comes out. If you have a spare minute i'd be interested to hear your take on 7DTD and virtual threads & core count. I know the gains in the tridents numbers above with the smt/off,6 core profile can largely be attributed to ryzen 3000's chiplet design, forcing 7dtd to use only cores on a single ccd + the higher oc. But even with the same clock speeds it performs significantly better with 4-6 cores and no hyper threading/smt. I tested it last on intel in a17 and setting manual cpu affinity to make 7dtd use only 4 real cores and assigning any background processes to the virtual threads resulted in +12% avg and +25% in the 1%lows, compared to just letting 7dtd use all 4 cores and 8 threads. Could it be the way hyper threading works? In the way even though task manager will see a 4 core with HT as an 8 core, it's still a 4 core and the virtual threads only work when there is a pause/break/gap in command queuing on the main thread, so if 2 process are time dependant on each other and they're both assigned to the 2 threads that are on the same core, if that main thread is maxed out the virtual thread has to wait for the few times the main thread is free/waiting on another process to finish. That's my best guess with a very limited understanding anyway 😜 I know scaling with more and more cores has been the white whale of game developers for years, but usually the answers are generally a very vague "it just hard" I probably won't understand 100% the actual reasons why, but i find this kind of stuff super interesting none the less.😃 Link to comment Share on other sites More sharing options...
Recommended Posts
Archived
This topic is now archived and is closed to further replies.