Performance Megathread

[Thinghunter]

Hey guys, now i have done my own testing on a big Galaxy. I can tell that the performance is quite good, especially the loading times.

My system specs:
- i9-9900k @4.7 ghz non oc
- 64gb ram @3200mhz
- rtx 2080ti
- game installed on samsung evo 970 M.2 SSD (as windows, steam & co are)

Game Settings:
- 1000 Systems
- 30 AI Empires
- All Fallen Empires/Caraveneers/Marauders
- 1x for pre FTL species/habitable planets/ftl lanes etc.
- standard settings for earlygame/midgame/endgabe/techs & co.
- Xeno compatibility activated
- difficulty captain, no advanced AI empires
- year 2407.04.20
- have roundabout 1400 pops; i am the farthest advanced
- every empire has ~400 pops on average -> should be about 14000 pops in total in the galaxy

I have tested the game 3 times and run in observer mode for exactly 4:00 minutes
- First test in system view, max zoomed out: i reached year 2409.02.10
- Second test in galaxy view, zoomed max out so that system ressources are visible: i reached year 2409.02.21 (11 days more)
- Third test in galaxy view, zoomed max out that fleet icons still visible: i reached year 2409.03.14 (basically 1 month more)

My CPU usage was constantly 25% -> 4 cores used, no hyperthreading
about 12.5 gb ram were used (Firefox open with some tabs, but no videos etc running, only the paradox forums)
GPU usage ~27% in all tests; 4.7 gb vram used

Loading times:
- 6 seconds for launcher from steam
- 14 seconds from launcher to game main menu
- 8 seconds from main menu to savegame


Summary:
Great performance improvements, especially very impressive loading times. Game totally playable, no lag till now- but i havent reached the very endgame yet so maybe endgame performance will suffer if pops/fleets get doubled again.

PS: no mods activated, every DLC activated besides Necroids

 

[Xmostera]

Hey guys, now i have done my own testing on a big Galaxy. I can tell that the performance is quite good, especially the loading times.

My system specs:
- i9-9900k @4.7 ghz non oc
- 64gb ram @3200mhz
- rtx 2080ti
- game installed on samsung evo 970 M.2 SSD (as windows, steam & co are)

Game Settings:
- 1000 Systems
- 30 AI Empires
- All Fallen Empires/Caraveneers/Marauders
- 1x for pre FTL species/habitable planets/ftl lanes etc.
- standard settings for earlygame/midgame/endgabe/techs & co.
- Xeno compatibility activated
- difficulty captain, no advanced AI empires
- year 2407.04.20
- have roundabout 1400 pops; i am the farthest advanced
- every empire has ~400 pops on average -> should be about 14000 pops in total in the galaxy

I have tested the game 3 times and run in observer mode for exactly 4:00 minutes
- First test in system view, max zoomed out: i reached year 2409.02.10
- Second test in galaxy view, zoomed max out so that system ressources are visible: i reached year 2409.02.21 (11 days more)
- Third test in galaxy view, zoomed max out that fleet icons still visible: i reached year 2409.03.14 (basically 1 month more)

My CPU usage was constantly 25% -> 4 cores used, no hyperthreading
about 12.5 gb ram were used (Firefox open with some tabs, but no videos etc running, only the paradox forums)
GPU usage ~27% in all tests; 4.7 gb vram used

Loading times:
- 6 seconds for launcher from steam
- 14 seconds from launcher to game main menu
- 8 seconds from main menu to savegame


Summary:
Great performance improvements, especially very impressive loading times. Game totally playable, no lag till now- but i havent reached the very endgame yet so maybe endgame performance will suffer if pops/fleets get doubled again.

PS: no mods activated, every DLC activated besides Necroids

is the CPU utilization result based on Single Core or overall 25%? It would be interesting if 4.7Ghz is being only utilized with 25%. That means they improved the game a lot compared to before. I am not sure whether the game is more memory speed sensitive/CPU. But the load time you are mentioning are mostly thanks to your samsung evo 970. Could I know how much seconds/minutes taken to run 1 month/12months after this year of 2409?

 

[Thinghunter]

is the CPU utilization result based on Single Core or overall 25%? It would be interesting if 4.7Ghz is being only utilized with 25%. That means they improved the game a lot compared to before. I am not sure whether the game is more memory speed sensitive/CPU. But the load time you are mentioning are mostly thanks to your samsung evo 970. Could I know how much seconds/minutes taken to run 1 month/12months after this year of 2409?

The CPU utilization benchmarked with MSI afterburner shows overall 25%. It doesnt mean that 4 cores are used at 100%, 4 cores are not used and that no hyperthreading is used. It means, that the CPU has enough headroom to better multithread (by stellaris) and multitask (e.g. by streaming on twitch)

About the loading times: I can say that it took considerably longer before the patch to load a game- perhaps 5 times that long- on the same pc.

About the time per month: Id say it takes about 130s to run a year, so about 11sec/month

 

[iZMXi]

Hello. If any gentlemen here are in possession of a Ryzen 5000 cpu and is not gpu bottenecked in stellaris in any way I would very much appreciate if they did a late game performance test by stopwatching how long it takes for x months or even a year to pass on average, preferrably over a few runs on Fastest, and have a 2-stick ram config and are willing to tell me what speeds this RAM runs at I would like to do some apples to apples comparison with Intel to determine whether it's a worthy upgrade to go from Intel 7th/8th/9th/10th gen up to Ryzen 5000, in other words whether it's one of those games AMD is leading by a few % or if it's one of the games they are leading by a large margin. For the good of the people who play stellaris that are considering whether or not to upgrade (including myself).

Please post your time results, ram speed, and the save file here if interested and I will run the same test on my 10700k to reach a conclusion about this. Make sure the save is late/heavy enough to actually be slowed down of course.

If you have a 4-stick ram config, I can't copy that for testing, but if you would be willing to test whether or not there is a difference between having single vs dual rank (running with 2 sticks instead of 4) to see if Stellaris is something that also benefits hugely from being able to simultaneously read and write to memory registers, that would also be very appreciated, but I don't expect anyone to actually go through that much effort.

You could also use my old performance-testing save from 2.8.0 but I am not 100% sure if it still works. Full vanilla. Thanks in advance.

My Ryzen 5 5600X @ allcore "overclock" of 4600MHz @ 1.275 volts with (2x16GB DR, 2 DIMMs, 4 ranks total) DDR4-3600 20-20-20-40 gets:
1 year in 3 minutes, 50 seconds. I ran it twice and got the same result.
I also tried it with RAM at 2133MHz and it took 4 minutes, 30 seconds.

 

[John_Titor]

Running a Ryzen 7 5800X and RTX 3080, both stock, and it shreds Stellaris now. Even late game on a huge map with a lot of empires it runs really well. Early game Fastest is blistering, i've never seen it that fast. I upgraded from a Ryzen 5 1600 and GTX 1070.

 

[GnoSIS]

Running a Ryzen 7 5800X and RTX 3080, both stock, and it shreds Stellaris now. Even late game on a huge map with a lot of empires it runs really well. Early game Fastest is blistering, i've never seen it that fast. I upgraded from a Ryzen 5 1600 and GTX 1070.

Can you provide a count of all pops, including those of AI empires?

I'm slated to upgrade in a few weeks or a month, provided things go smooth with the PC component/laptop market.

 

[Thinghunter]

Well i have a 9900k & 2080ti and i currently have acceptable speed in lategame as i mentioned earlier. If you use a 5800x & a rtx 3080, you may get 10% to 20% faster gameplay than i do. Depending on how your current system is, i think an upgrade is not worth it if you get only a 30% faster gameplay. This means i would upgrade to play Stellaris in my usecase (max galaxies, max AI empires, mods) if i have less than a 8700k, 9900k, 3600x, 5600x etc. as i strongly recommend to have 6 cores/12 threads for playing Stellaris on these settings. If you play on smaller maps with no mods, 4 cpu cores & 8 threads should be okay if the cpu is a 4 core/8 thread intel (6700k or newer) or a 4 core/8 thread zen 2 amd cpu (3300x).
If i had to buy a CPU for Stellaris for my usecase, i would consider a 5900x or a 5950x despite the significant higher cost than a 5600x & 5800x as i hope that stellaris will be able to use more than 8 cores in the endgame as the game gets patched in the upcoming years. Again, if i would play stellaris on a smaller map with less AI empires, a 5600x or 5800x would surely be very sufficient for that, but i personally would not spend a lot of money on a pc upgrade if i only get the best experience for a short time (if i assume that Stellaris can utilize more cores).

 

[ltkAlpha]

Well i have a 9900k & 2080ti and i currently have acceptable speed in lategame as i mentioned earlier. If you use a 5800x & a rtx 3080, you may get 10% to 20% faster gameplay than i do. Depending on how your current system is, i think an upgrade is not worth it if you get only a 30% faster gameplay. This means i would upgrade to play Stellaris in my usecase (max galaxies, max AI empires, mods) if i have less than a 8700k, 9900k, 3600x, 5600x etc. as i strongly recommend to have 6 cores/12 threads for playing Stellaris on these settings. If you play on smaller maps with no mods, 4 cpu cores & 8 threads should be okay if the cpu is a 4 core/8 thread intel (6700k or newer) or a 4 core/8 thread zen 2 amd cpu (3300x).
If i had to buy a CPU for Stellaris for my usecase, i would consider a 5900x or a 5950x despite the significant higher cost than a 5600x & 5800x as i hope that stellaris will be able to use more than 8 cores in the endgame as the game gets patched in the upcoming years. Again, if i would play stellaris on a smaller map with less AI empires, a 5600x or 5800x would surely be very sufficient for that, but i personally would not spend a lot of money on a pc upgrade if i only get the best experience for a short time (if i assume that Stellaris can utilize more cores).

I'm new to Stellaris. It was one of the many games I was mostly collecting DLC for without playing for the past few years. I started playing after my recent upgrade to a 5800X. I'm also running 32GB of fairly well tuned RAM (3600 MT/s, 15-15-15-31), a water cooled 1080 Ti and a good Gen3 NVMe SSD. I found this thread when looking for Zen related tips for performance optimization, because as I progressed through to the late game I noticed significant slow downs. Having a few more hours to look into this since I started looking around I thought I'd share a few observations.

The performance of Stellaris is CPU bound. You can see that by looking at the CPU and GPU utilization numbers in the Afterburner overlay in game. In the late game there is always one thread that sits at 98+% utilization. There's usually a secondary thread hovering around 30%, but whether or not it's connected to the game or the YouTube videos I'm playing in the background I cannot say. The rest of the threads do little - 5-15% - but there may be some multi-threading going on.

Worse, at present Stellaris is utterly bottlenecked by that one thread, which appears to share the state-of-the-game engine and graphics rendering logic. You can test that by loading a late-game save and pausing and unpausing the game at Fastest - when I pause the game my 1080 Ti immediately bumps into the frame cap of 142 (which I've set to keep refresh rate within the G-Sync window of my display and avoid screen tearing) at around 40-70% GPU utilization. Worryingly, the main thread keeps utilization of its logical processor at around 95% even when the game's paused (!). The moment you unpause GPU utilization plummets to sometimes below 10%, presumably because rendering has to wait to receive a game state before it can render it. I suspect that if that weren't the case most most modern graphics cards would be fine for the game except in perhaps the largest battles.

Unfortunately, what I've described above is a bit of a worst case from a player perspective. Given the obvious single thread dependence of the game, unless you can buy a CPU with cores that are twice as powerful (which won't be soon and may not matter as much if one's looking for fluid performance) or unless Paradox decide to re-architect their game to be much more multi-threaded, we aren't going to see substantial change to performance at this point in the game's life cycle. Whether or not something like that is on the table only Paradox can say. Given how complex such an exercise appears to be (look up the GDC talk Multithreading the Entire Destiny) I have my doubts.

 

[GnoSIS]

I'm new to Stellaris. It was one of the many games I was mostly collecting DLC for without playing for the past few years. I started playing after my recent upgrade to a 5800X. I'm also running 32GB of fairly well tuned RAM (3600 MT/s, 15-15-15-31), a water cooled 1080 Ti and a good Gen3 NVMe SSD. I found this thread when looking for Zen related tips for performance optimization, because as I progressed through to the late game I noticed significant slow downs. Having a few more hours to look into this since I started looking around I thought I'd share a few observations.

The performance of Stellaris is CPU bound. You can see that by looking at the CPU and GPU utilization numbers in the Afterburner overlay in game. In the late game there is always one thread that sits at 98+% utilization. There's usually a secondary thread hovering around 30%, but whether or not it's connected to the game or the YouTube videos I'm playing in the background I cannot say. The rest of the threads do little - 5-15% - but there may be some multi-threading going on.

Worse, at present Stellaris is utterly bottlenecked by that one thread, which appears to share the state-of-the-game engine and graphics rendering logic. You can test that by loading a late-game save and pausing and unpausing the game at Fastest - when I pause the game my 1080 Ti immediately bumps into the frame cap of 142 (which I've set to keep refresh rate within the G-Sync window of my display and avoid screen tearing) at around 40-70% GPU utilization. Worryingly, the main thread keeps utilization of its logical processor at around 95% even when the game's paused (!). The moment you unpause GPU utilization plummets to sometimes below 10%, presumably because rendering has to wait to receive a game state before it can render it. I suspect that if that weren't the case most most modern graphics cards would be fine for the game except in perhaps the largest battles.

Unfortunately, what I've described above is a bit of a worst case from a player perspective. Given the obvious single thread dependence of the game, unless you can buy a CPU with cores that are twice as powerful (which won't be soon and may not matter as much if one's looking for fluid performance) or unless Paradox decide to re-architect their game to be much more multi-threaded, we aren't going to see substantial change to performance at this point in the game's life cycle. Whether or not something like that is on the table only Paradox can say. Given how complex such an exercise appears to be (look up the GDC talk Multithreading the Entire Destiny) I have my doubts.

Here, we've been discussing this to death, in these 77 pages, years ago, since the Megacorp economy update.

We, can't expect a full engine redesign. However with the next patch, they are scaling population back, and they change pop growth to be based on your empires existing total count, and thus have a potential "asymptotic" max pop effect. This will potentially allow us to play the endgame normally. But don't expect to play for 1000 years, with huge pop counts.

You have a top end rig, so you should be able to play in the most confortable way. That is what the game is and will be for years on. Only expect major engineering/engine progress in a future sequel, if they ever make it, and anything else, will be huge news for everyone here.

The whole multithreading bandwagon will be going on and on in gamedev, for many years, because almost 100% of games out there are indeed in effect singlethreaded for game state, with a few exceptions. And now that an 8 cores/16 thread CPU, or a 16 core/ 32 thread one is common place, people will be very grumpy everywhere.

P.S. I'm getting the same rig as yours, soon.

P.S. 2: You're running your DIMMS slower than 3600Mhz: Timings should be: 18-22-22-42, "NB" frequency 1800Mhz, Voltage 1.35V. Check your BIOS.
15-15-15-31 timings are for 2133MHz total, 1064.5MHz per channel. For example see: https://www.corsair.com/uk/en/Categ...RGB-Black/p/CMW16GX4M2D3600C18#tab-tech-specs. There is no standard or DIMMS with those timings for 3600Mhz.

 

[ltkAlpha]

Here, we've been discussing this to death, in these 77 pages, years ago, since the Megacorp economy update.

We, can't expect a full engine redesign. However with the next patch, they are scaling population back, and they change pop growth to be based on your empires existing total count, and thus have a potential "asymptotic" max pop effect. This will potentially allow us to play the endgame normally. But don't expect to play for 1000 years, with huge pop counts.

You have a top end rig, so you should be able to play in the most confortable way. That is what the game is and will be for years on. Only expect major engineering/engine progress in a future sequel, if they ever make it, and anything else, will be huge news for everyone here.

The whole multithreading bandwagon will be going on and on in gamedev, for many years, because almost 100% of games out there are indeed in effect singlethreaded for game state, with a few exceptions. And now that an 8 cores/16 thread CPU, or a 16 core/ 32 thread one is common place, people will be very grumpy everywhere.

P.S. I'm getting the same rig as yours, soon.

P.S. 2: You're running your DIMMS slower than 3600Mhz: Timings should be: 18-22-22-42, "NB" frequency 1800Mhz, Voltage 1.35V. Check your BIOS.
15-15-15-31 timings are for 2133MHz total, 1064.5MHz per channel. For example see: [link] There is no standard or DIMMS with those timings for 3600Mhz.

Good to hear a workaround is being worked on.

On game state being single threaded for most games - I wonder if one could split AI actors onto separate threads and have those communicate with the main game state thread. I'm sure it wouldn't be a trivial challenge (if it's doable at all), but I'd be curious to hear a dev who's encountered something similar speak on the topic.

Re. my rig - I was running an OCed 8700K until the holidays; the whole point of the upgrade was to move to an RTX 3080 for Cyberpunk as there was an issue with my CPU/board combo whereby it wouldn't run more than PCIe x2 on the x16 slot, so I was limited to x8 on the secondary PCIe slot and I didn't want to limit my flashy new 3080. I decided to go up 2 cores if I'm upgrading. Well, we all know what happened on the GPU front. Fortunately, since I won't be buying Cyberpunk before they release a GotY edition I have no pressing need for a GPU update + I have literally dozens of older games to enjoy. I'm happy with the 5800X - it's fast and it overclocks itself (although temps and voltages are scary without prior experience with the architecture)!

Re. the RAM speed - I'm not, don't worry. When I mentioned tuned, I meant hand tuned and performance tested by me with the help of DRAM Calculator for Ryzen (which roughly confirmed the primary timings I had already established on my Intel board). 15-15-15-30 settings aren't much of a stretch for Samsung B-Die. I'm running 1.39V and there are kits out there running 1.45V as XMP. My particular Kingston kits were meant to run an XMP of 16-18-18-36 @ 3466 @ 1.35V, so second rate B-Die. I'm running tRAS below the recommended setting for Fast because I tested 30 and it was slower than 31 (which was faster than 32). The settings are MemTestPro stable up to 800% (where I stopped the test the next day) and y-cruncher stable for 24 hours (again, where I stopped the test). The RAM was fine running 16-16-16-32 @ 3800, but the CPU was not and was throwing WHEA errors by the hundreds (without me noticing, initially), so I'm going to wait a year for BIOS/AMD AGESA to mature and give 1900 another try.

 

[GnoSIS]

Good to hear a workaround is being worked on.

On game state being single threaded for most games - I wonder if one could split AI actors onto separate threads and have those communicate with the main game state thread. I'm sure it wouldn't be a trivial challenge (if it's doable at all), but I'd be curious to hear a dev who's encountered something similar speak on the topic.

The problem with all state, is mutability. If you have AI actors making decisions, at some point, after they decide what to do they must interact with the state and change it. Here lies the problem: You can't have that. You can't even read state, while you make changes to it, and there is no easy and clean workaround or any cowboy/spaggeti code solution- full stop!! And by extension, you can't render the state, while making changes to it, because rendering requires reading the state.

Same for all commercial engines. Unity and Unreal, All of their included state frameworks (Entity Component and friends) are strictly single threaded.

Ok, cool about your ram setup!

 

[Jamey]

Same for all commercial engines. Unity and Unreal, All of their included state frameworks (Entity Component and friends) are strictly single threaded.

There are games designed from the ground up to be multithreaded which are quite good at it. Space Engineers comes to mind as a game that can max out every one of my cores if I go overboard within.

The problem many games have is that they are not designed for multithreading (and, frankly, most games just don’t need it much).

 

[GnoSIS]

There are games designed from the ground up to be multithreaded which are quite good at it. Space Engineers comes to mind as a game that can max out every one of my cores if I go overboard within.

The problem many games have is that they are not designed for multithreading (and, frankly, most games just don’t need it much).

Yes, but it's done in a non standard way, and each developer re-invents the wheel.
The issue with grand strategy and other similar games is that there's huge state, that gets touched very often.

Solutions are:
1. Spend milions in engineering to do it correctly.
2. Reduce state/simplify game model. (Make pops abstract.. or reduce pops)
3. Reduce state access frequency. I.e. do pop checks once a year.

 

[ltkAlpha]

Yes, but it's done in a non standard way, and each developer re-invents the wheel.
The issue with grand strategy and other similar games is that there's huge state, that gets touched very often.

Solutions are:
1. Spend milions in engineering to do it correctly.
2. Reduce state/simplify game model. (Make pops abstract.. or reduce pops)
3. Reduce state access frequency. I.e. do pop checks once a year.

Isn't looking for those solutions something that just comes with the territory, though? I'm sure one needs to make tradeoffs to keep a project commercially viable and Bungie had significantly more resources at their disposal (since I mentioned their GDC talk on Destiny earlier), but the result must've been unsatisfying (edit: for the devs, at launch), both practically in terms of performance and user experience and conceptually, since running into the limitations of the main thread is just a matter of time. I mean, here we are...

 

[GnoSIS]

Isn't looking for those solutions something that just comes with the territory, though? I'm sure one needs to make tradeoffs to keep a project commercially viable and Bungie had significantly more resources at their disposal (since I mentioned their GDC talk on Destiny earlier), but the result must've been unsatisfying, both practically in terms of performance and user experience and conceptually, since running into the limitations of the main thread is just a matter of time. I mean, here we are...

Yes and no,

Since gaming exploded, it has been the driving force for CPU innovation. Historically, companies waited for the next gen CPU that gave them X more single threaded performance, and it worked for decades. Same for stellaris, with a new 2021 system, you play ok.

The problem with the developer pool: Experience in the industry is very low as far as multithreading goes, and most developers know the tried and true "single threaded techniques" more or less. Many developers don't choose to work in gamedev, due to long hours, low salaries etc... If you are a code wizard, you go out and ask for a wizards salary. But I'm generalizing here, I can expand this paragraph into an entire book - and I'm not trying to put blame.

The problem with the publishers: Existing tools and proprietary internal engines, that worked in past titles are used with small enhancements in new projects, as was the case with stellaris. But an old horse can go that far ahead. Destiny was not coded from scratch, and as with stellaris, you keep adding features and push the engine beyond its limits and this is the result.

Basically the industry has technical debt to unload from the mainthread that goes back 20 years or more, but you can't deliver a game on time and dev budget doing that. If you check the data for it, you will discover that the budget for actual coding/engine dev is very small, most studios throw millions to have Keanu Reeves, have huge motion capture studios and hire hordes of 3D artists and animators instead, and their managers will tell you that those things sell.

 

[grommile]

their managers will tell you that those things sell.

they'll tell you that.

what they mean is "this is the only thing we know how to sell".

a good sales/marketing team can sell sand to the saudis and defective lab glassware to companies making ornamental concrete castings.

 

[Jamey]

The issue with grand strategy and other similar games is that there's huge state, that gets touched very often.

Which is a hard, but solvable problem.

Solutions are:
1. Spend milions in engineering to do it correctly.
2. Reduce state/simplify game model. (Make pops abstract.. or reduce pops)
3. Reduce state access frequency. I.e. do pop checks once a year.

I‘d argue that 2 and 3 are not incorrect ways of optimizing things, as long as it doesn’t negatively impact player experience.

Multithreaded coding is hard, especially if you are working with a legacy code base. There’s a reason that Imperator and CK3 run so fast relative to the older games. Paradox clearly made it a priority, probably due to the valid complaints about their older games running slowly.

I’m not sure I’ll ever play Stellaris again. It’s been over two years since I had fun playing it, and the reason why is the mid to late game slowdown introduced with planet and pops rework. I am curious if they’re going to finally fix performance, but I’m also not sure that I’ll have the ability to start another play through because I’m not sure that I’ll believe that they actually have improved performance this time.

 

[James Fire]

a good sales/marketing team can sell sand to the saudis

Kinda offtopic, but very interesting, is that lots of places actually do sell sand to the Saudis.

 

[WarriorofMODS]

Which is a hard, but solvable problem.


I‘d argue that 2 and 3 are not incorrect ways of optimizing things, as long as it doesn’t negatively impact player experience.

Multithreaded coding is hard, especially if you are working with a legacy code base. There’s a reason that Imperator and CK3 run so fast relative to the older games. Paradox clearly made it a priority, probably due to the valid complaints about their older games running slowly.

I’m not sure I’ll ever play Stellaris again. It’s been over two years since I had fun playing it, and the reason why is the mid to late game slowdown introduced with planet and pops rework. I am curious if they’re going to finally fix performance, but I’m also not sure that I’ll have the ability to start another play through because I’m not sure that I’ll believe that they actually have improved performance this time.

There's also the fact that the sales of Stellaris 2 will suffer due to it's terrible reputation.

 

[GnoSIS]

they'll tell you that.

what they mean is "this is the only thing we know how to sell".

a good sales/marketing team can sell sand to the saudis and defective lab glassware to companies making ornamental concrete castings.

Yes,

but as much as developer salaries and quality is concerned, this is the current culture for management.

So while you're right, it's the environment we're in that counts. Exceptions exist: Wube software - Factorio.