The hallmark and symbol of the modern world is advertising, above technology and industry. Imagine spending billions, building up the "premium" product.
"My branded sand is better than the sand you have already. It also has a nice flower scent!"
- 1
Performance Megathread
- Thread starter Ashantai
- Start date
-
We have updated our Community Code of Conduct. Please read through the new rules for the forum that are an integral part of Paradox Interactive’s User Agreement.
"My branded sand is better than the sand you have already. It also has a nice flower scent!"
I'm back again to do some more hardware testing... I'm still kinda salty my old results have been rendered obsolete by the game updates but even happier to see the game running so much better. Anyway, Intel 11th gen is right around the corner later this month and it's bringing the first major core redesign to Intel's desktop platform in a very long time... It is essentially this but backported to 14nm and benchmarks so far suggest is nearly if not entirely unchanged. Still, I want to do some apples to apples comparison with 10th gen to get some idea as to whether it is a significant improvement to late game at all over Skylake (6th-10th gen Intel, which all perform nearly the same in this game).
My overall goals
1. compare jedec 2133 vs tightened down ddr4-3600 performance on both the i7-10700k and i7-11700k at the same clock/ring speeds
2. desync vs 1:1 vs 1:2 FCLK performance on 11th gen (new feature)
3. see how Zen 3 stacks up in this game compared to both architectures, and whether it is king for this game
1 and 2 will take a lot of time testing to get all the data for, but I will need some help with #3. If anybody here has a Zen 3 cpu (5600x, 5800x, 5900x, or 5950x) and wants to be a bro, please have or make a late game vanilla save that is slowing down noticeably already, make a copy of that save file for me as is when you loaded in, and then go to the galaxy map as observer and let exactly one year go by on Fastest starting from where the save loads in, hit pause again on that very day. Use a stopwatch and sync it with the unpause/pause to get an exact measurement of how long this took. I personally suspect at this point that AMD might actually be king for Stellaris with its monstrous caches and single thread performance that exceeds both Intel 10th and 11th gen, and that it will get through this year faster in spite of its higher memory latency, but I would really need to compare directly to be sure. If you want to do multiple runs or share your memory configuration/timings etc. that would also be very helpful, but at this point any data at all about Zen 3's Stellaris performance would be better than what I presently have to work with, which is nothing.
Just make sure you aren't GPU bottlenecked somehow, and that you post the save file you copied before hitting unpause to stopwatch it here along with your recorded times so that I can use the same save file to conduct the test in the same way. If you don't have a vanilla save and still want to help, my usual approach to make one would be 5x habitable max ai galaxy with 1000 stars, 0.25x tech, ironman off, fast_forward 72000 (200 years) and let it progress. Could do less than 200 years, I just like a really unrealistically intense late game scenario because the longer the year takes, the higher the resolution and accuracy of the results. If you want to go even further than 200 that's entirely fine, too. Also, there's no need to stopwatch the fast_forward, I'm not using that data.
Let me know what version you're using too, ideally it should be either the latest or the beta. If you have a Zen 2 chip (3600/3700x, etc.) your data can also be useful to me, I'm just more interested in Zen 3 right now because it is the new king of gaming.
I'll proceed with 10th vs 11th gen regardless of whether or not I can get Ryzen results, but since 11th gen is still not out for about 4 weeks, and I don't know how long after that it will take for me to get my hands on it, I'll give the next 2-3 weeks at least for some cool AMD owner to help me make these tests more useful.
By the end of all this we should have a pretty clear idea of whether AMD or Intel is the ideal choice for Stellaris, and whether memory timings still matters as much with these newer architectures that have deeper/wider pipelines and better cache/branch predictors to help them iterate faster through all those nasty jobs, pops, ships, planets, etc.
My overall goals
1. compare jedec 2133 vs tightened down ddr4-3600 performance on both the i7-10700k and i7-11700k at the same clock/ring speeds
2. desync vs 1:1 vs 1:2 FCLK performance on 11th gen (new feature)
3. see how Zen 3 stacks up in this game compared to both architectures, and whether it is king for this game
1 and 2 will take a lot of time testing to get all the data for, but I will need some help with #3. If anybody here has a Zen 3 cpu (5600x, 5800x, 5900x, or 5950x) and wants to be a bro, please have or make a late game vanilla save that is slowing down noticeably already, make a copy of that save file for me as is when you loaded in, and then go to the galaxy map as observer and let exactly one year go by on Fastest starting from where the save loads in, hit pause again on that very day. Use a stopwatch and sync it with the unpause/pause to get an exact measurement of how long this took. I personally suspect at this point that AMD might actually be king for Stellaris with its monstrous caches and single thread performance that exceeds both Intel 10th and 11th gen, and that it will get through this year faster in spite of its higher memory latency, but I would really need to compare directly to be sure. If you want to do multiple runs or share your memory configuration/timings etc. that would also be very helpful, but at this point any data at all about Zen 3's Stellaris performance would be better than what I presently have to work with, which is nothing.
Just make sure you aren't GPU bottlenecked somehow, and that you post the save file you copied before hitting unpause to stopwatch it here along with your recorded times so that I can use the same save file to conduct the test in the same way. If you don't have a vanilla save and still want to help, my usual approach to make one would be 5x habitable max ai galaxy with 1000 stars, 0.25x tech, ironman off, fast_forward 72000 (200 years) and let it progress. Could do less than 200 years, I just like a really unrealistically intense late game scenario because the longer the year takes, the higher the resolution and accuracy of the results. If you want to go even further than 200 that's entirely fine, too. Also, there's no need to stopwatch the fast_forward, I'm not using that data.
Let me know what version you're using too, ideally it should be either the latest or the beta. If you have a Zen 2 chip (3600/3700x, etc.) your data can also be useful to me, I'm just more interested in Zen 3 right now because it is the new king of gaming.
I'll proceed with 10th vs 11th gen regardless of whether or not I can get Ryzen results, but since 11th gen is still not out for about 4 weeks, and I don't know how long after that it will take for me to get my hands on it, I'll give the next 2-3 weeks at least for some cool AMD owner to help me make these tests more useful.
By the end of all this we should have a pretty clear idea of whether AMD or Intel is the ideal choice for Stellaris, and whether memory timings still matters as much with these newer architectures that have deeper/wider pipelines and better cache/branch predictors to help them iterate faster through all those nasty jobs, pops, ships, planets, etc.
Last edited:
@Riince2 New gaming laptop arrives tomorrow, Ryzen 7 5800H and RTX 3070 mobile. It will take me time to setup and benchmark it.
Not what you were expecting, but close!
If you can post/share a reference vanilla game save, it would be cool.
Not what you were expecting, but close!
If you can post/share a reference vanilla game save, it would be cool.
- 1
They should stop adding cores and start making memory access faster. All those CPUs are just sitting there, and even if game engines were written to use them, the bottleneck would be memory access. Try feeding 8C 16T with a current gen bus. We needed DDR6 Ram DIMMS 5 years ago.
- 2
yeah, dram access has been the pervasive issue in cpu's since the 90's, and it's really only gotten worse since then, that's for sure. Almost every cpu design decision aimed at performance is based around increasing cache hits and reducing the penalty of misses so that the small percentage of instructions that need to be fetched from nightmarishly slow to access memory are as bearable as possible. Always an issue, but the way engineers have compensated for it is brilliant. "Effective" latency is going down all the time because of it, even if RAM itself is barely advancing. Of all human accomplishments, microprocessors are the one most able to be likened to magic...
Not exactly what I was expecting... but it's still a Zen 3 part! I didn't expect anyone to really step up so soon but I'll post a save either later today or tomorrow, it takes some time to fast forward. Thanks. Anyone else is welcome to download it too when I post it and share their results, no such thing as useless data...
Make sure the game uses your 3070, laptops with IGPU's and DGPU's can get a little fucky wucky with which one they use by default.
- 1
Yeah no worries, been aware of that for almost a decade, including the hundreds of PC I deal with everyday.
I've been playing on a 6 year old premium gaming laptop, and was planning for an upgrade since last year. I'm not early to step up, I'm retiring the "old general"
Perhaps this, along with the next update, will make me a less grumpy Stellaris player!
As much as I love large caches, if our miniature gaming universes are several GB large and we process all of that more or less each frame, no amout of cache will help.
We must start thinking about new layers i.e. adding a per core 1GB level 4 & 5 cache.
civ VI is the closest game to stellaris i've seen compared in reviews (it's 90% third/first person snorefest games) and Zen 3 is horrifyingly good in it. I have to assume it's mostly the cache. It may not all fit in the cache but even a notably larger % of it fitting should drastically increase cache hits compared to a part with less cache... and improve effective latency accordingly.
No idea how Stellaris will look though, it's still pretty different from Civ. Maybe I'll graph out my results 10th vs 11th vs zen 3 so it's pretty to look at.
Attachments
That graph is for FPS, not turn times.
I think the trick with Zen 3 is the per chiplet and per core cache revamp that makes such a big difference.
Good point, the difference is much smaller in turn times. Zen 3 still shreddin'. I do hope the 11700k surprises us.
Attachments
- 1
I just got a 5800x last week, and I have an RTX 3080 founders GFX card. I run the game at 1440p, I'll look into benchmarking per your criteria at the weekend.
I don't think it'll impact much (if at all) performance for Stellaris, but I'll benchmark again with the same conditions when re-sizeable BAR get's enabled by NVIDIA at the end of the month.
this took a little while to get to since i dont have access to the ideal system for benchmarking at the moment, but this should work.
no dlc or mods... 2.8.1 checksum dfce. jan 11th to jan 11th, good luck
i probably should have set habitable worlds higher but it's still close to 20k pops. Slowdown should be happening pretty well.
no dlc or mods... 2.8.1 checksum dfce. jan 11th to jan 11th, good luck
i probably should have set habitable worlds higher but it's still close to 20k pops. Slowdown should be happening pretty well.
Last edited:
My new Laptop Has Arrived!!!
Store/Model Link: https://uk.store.asus.com/gaming/no...board-windows-10-pro-grey-fa506qr-az001t.html(model is de-listed now, you know how it is with the current shortage. It was an odyssey to get this one! Been trying since january...)
For anyone looking to buy something similar, I was really impressed with the thermals. Asus did a great job on this model/line - no overheat & no burning keyboard/surfaces, even while the machine is at combined 100% CPU and GPU load. Only shame is that they didn't push the DDR specs to 3600MHz.
1. Obligatory 3dmark result for firestorm:
https://www.3dmark.com/fs/250647282. Some metrics on the above save, that has 20170 pops:
A. 1 year passage time for year 2402 (one year to warm up) in seconds (includes heavy tick at the start/end of year): 143.65
Average time for a month = 11.97 seconds i.e. 12 seconds per month.
B. CPU core activity: 1 thread at full speed, another one about 40-50% the other 14, almost idle. The chip just sits there, not even raising temperature and not sweating at all.
C. GPU activity: irrelevant, most of the time GPUz reports prefCap: Idle, gpu load in galaxy map and in normal capital system (earth) is less than 45% with reduced GPU clock speed. Settings are all set to max, multisampling at x8, monitor refresh and game set at 144Hz.
Conclusion: An order of magnitude faster compared to my previous machine, which was and still is above recomended specifications for this product.
**This laptop is overkill for Stellaris in specs, you shouldn't need something like this to play**.
Last edited:
- 1
- 1
that is surprisingly fast, i probably should have made the save even heavier or gone with 5x habitable instead of 2x. I'll still use it I guess. At least there is a little bit of slowdown but the resolution isn't as high as I'd have liked to see considering a start of game year is about 60-70 seconds on fastest...Store/Model Link: https://uk.store.asus.com/gaming/no...board-windows-10-pro-grey-fa506qr-az001t.html
(model is de-listed now, you know how it is with the current shortage. It was an odyssey to get this one! Been trying since january...)
For anyone looking to buy something similar, I was really impressed with the thermals. Asus did a great job on this model/line - no overheat & no burning keyboard/surfaces, even while the machine is at combined 100% CPU and GPU load. Only shame is that they didn't push the DDR specs to 3600MHz.
1. Obligatory 3dmark result for firestorm:
https://www.3dmark.com/fs/25064728
2. Some metrics on the above save, that has 20170 pops:
A. 1 year passage time for year 2402 (one year to warm up) in seconds (includes heavy tick at the start/end of year): 143.65
Average time for a month = 11.97 seconds i.e. 12 seconds per month.
B. CPU core activity: 1 thread at full speed, another one about 40-50% the other 14, almost idle. The chip just sits there, not even raising temperature and not sweating at all.
C. GPU activity: irrelevant, most of the time GPUz reports prefCap: Idle, gpu load in galaxy map and in normal capital system (earth) is less than 45%. Settings are all set to max, multisampling at x8, monitor refresh and game set at 144Hz.
Conclusion: An order of magnitude faster compared to my previous machine, which was and still is above recomended specifications for this product.
**This laptop is overkill for Stellaris in specs, you shouldn't need something like this to play**.
- 1
Start of game year, is bananas fast - unplayable. These speed settings have no metrics, so they can go as fast as they like.
- 1
Here's a save with 31k pops. (no mods, excluding some admin to count pops and some visual ones)
Max habitability, max empires, advanced start(all), agressive, grand admiral, huge.
You would have thought that it would take double the pops to halve the year time. Unfortunaly that's not the case. At 47% more pops, the time doubles.
One year of the attached save is 273.76 seconds, 22.81 per month.
And check the date! only 2359!
So should I wait for the technology to advance 6 more years, and buy a new laptop again then to play?
This is why pop scaling and inverse logarithmic pop growth is sorely needed. At a 33% pop scale, I would play this save at double the speed!
- 3
I'm adding a further datapoint.
This is a bit unfair/unrealistic, beacause most AI are in a WIH alliance and there's no wars going on. So the save runs a bit faster compared to the others, because with so many AI empires it's impossible to find a situation where everyone is at peace.
Year: 2419, continuation of above save.
Pops: 51300
Year to year seconds: 381.54 seconds (6 mins, 21.54 secs)
Average monthy time: 31.79, that's 33% slower compared to 31k pops
I'll try to post 2 more datapoints, at 61k and 71k, but the WIH and crisis are starting now, not sure if galactic pops will manage to survive to such levels.
(check my post above for specs if you're seeing just this post)
This is a bit unfair/unrealistic, beacause most AI are in a WIH alliance and there's no wars going on. So the save runs a bit faster compared to the others, because with so many AI empires it's impossible to find a situation where everyone is at peace.
Year: 2419, continuation of above save.
Pops: 51300
Year to year seconds: 381.54 seconds (6 mins, 21.54 secs)
Average monthy time: 31.79, that's 33% slower compared to 31k pops
I'll try to post 2 more datapoints, at 61k and 71k, but the WIH and crisis are starting now, not sure if galactic pops will manage to survive to such levels.
(check my post above for specs if you're seeing just this post)
- 1
- 1
For anyone experiencing lag hell in 2400 here is a good command for you . ticks_per_turn = 2. Use that in 2400 and it will be going fater then fastest in 2200. But at the cost of the game running at 10fps or so. Warning it will go so fast that you will have to pause to do everything. Edited because the fourms turn my emodicons into cringe emojis
Last edited:
Sounds like there's a quadratic time algorithm in there somewhere (time = pops squared). Something that compares each pop to all other pops, or all others in the same empire at least, would do that.
Job assignments and pop/job reviews on all planets. We found about this almost 1 year ago.
- 2