Best CPU available: AMD Ryzen 9 7950X3D?

[SoShootMe]

It's impossible for the OS to reliably figure out that the best thing to do is restrict a process/some thread(s) to the cores with extra cache available.

I wouldn't go that far. Maybe nobody does it, but I'm not sure even about that - - AIX would notice when pairs of processes interfered with each other and schedule them into different cache domains, it's not hard to imagine an OS being taught to set up the performance counters so as to track cache-miss-per-retirement stats and schedule processes where they get the biggest wins.

Yes; this is similar to what I described (not really relevant here, but I'm quite interested in the specific example you gave - can you give any pointers to more information?). I didn't say it is impossible to do, only that it is impossible for it to be reliable, ie to have any guarantee that the OS will choose the best-performing solution in all cases, which would remove the need for "manual tuning".

To look at it the other way around: no matter the scheduler improvements that may be made, it remains possible that it won't be optimal for Factorio, or any other specific case. Perhaps it will get close enough, perhaps not.

[datarza]

I am very happy to hear that in order to play comfortably in a game that was released in 2016, after 7 years, you need a threadripper.

[jodokus31]

I am very happy to hear that in order to play comfortably in a game that was released in 2016, after 7 years, you need a threadripper.

A Threadripper is not the way to go. You need a CPU with high single core performance. And those X3D processor from AMD also have a huge L3 cache, which also seems to be optimal for factorio.
All in all, you can play this game on a potato, but you just can't build huge factories without slow down.

[quyxkh]

And just to put some scale on "huge", six years ago, in 2017, you could build past 4kspm on a i3 4170, with the 0.15 engine. at 50fps.

[datarza]

I am very happy to hear that in order to play comfortably in a game that was released in 2016, after 7 years, you need a threadripper.

A Threadripper is not the way to go.

Probably not, probably yes, want to see the real test.

You need a CPU with high single core performance.

I do not think so, proof:

Untitled.png (114.55 KiB) Viewed 4582 times

and second proof for single core:

Untitled.png (260.31 KiB) Viewed 4577 times

And those X3D processor from AMD also have a huge L3 cache, which also seems to be optimal for factorio.

Not optimal, but it looks like a maximum for today, because of CPU cache. I believe there is no good processor for Factorio and never be, till they (owners) fix the inserters (probably never).

[jodokus31]

I am very happy to hear that in order to play comfortably in a game that was released in 2016, after 7 years, you need a threadripper.

A Threadripper is not the way to go.

Probably not, probably yes, want to see the real test.

You need a CPU with high single core performance.

I do not think so, proof:

Untitled.png

And those X3D processor from AMD also have a huge L3 cache, which also seems to be optimal for factorio.

Not optimal, but it looks like a maximum for today, because of CPU cache. I believe there is no good processor for Factorio and never be, till they (owners) fix the inserters (probably never).

Ok, I should have said, you need a CPU with high single core performance and huge memory bandwidth. Memory is probably even more important than CPU power.
Those X3D processors have the huge L3 cache and help with memory bandwidth.
Maybe not absolutely the optimal but relative to others, currently.

Multicore performance is also important, because a lot of things got parallelized, if I understood correctly, but some thread will always be the bottleneck.

[datarza]

Ok, I should have said, you need a CPU with high single core performance

Second time repeat: CPU frequency in Factorio is not an issue

[jodokus31]

Ok, I should have said, you need a CPU with high single core performance

Second time repeat: CPU frequency in Factorio is not an issue

I don't understand, what you are trying to say.

My statement is, that you need a CPU with high single core performance and a high memory bandwidth. That is, what I heard over the years dealing with factorio. My experience is also, that you have a hard time to parallelize a single application to use all core evenly, unless you have tasks, which can be parallelized evenly. The moment, you have a main thread like factorio or most other games, you can't really fully parallelize everything evenly, because you have to synchronize it back and one thread will scratch on the limit of the single core somewhere.

You said, a CPU with high single core performance is not needed and show some screenshots, which display, that all cores are used somehow but not to max. but UPS is still tanked. And you say it's a proof against my statement. Clearly, some bottleneck is going on in your scenario, but you think it's the amount of CPU cores or what else? I'm sorry, I can't see it as proof, I would rather say, that your memory bandwidth is problematic or something else.

Btw.: I never said something about CPU frequency, but obviously different CPU Architectures like AMD and Intel have different frequency for similar performance. f.e the 5800X3D has a lower frequency than the 5800X, but it still has more gaming performance.
That said, higher frequency for the same CPU obviously is always more performant, whether you need it or not in case of other bottlenecks.

[datarza]

Ok, I should have said, you need a CPU with high single core performance

Second time repeat: CPU frequency in Factorio is not an issue

I don't understand, what you are trying to say.

you need a CPU with high single core performance

No, I do not need it. And sorry, I did not read about your experience, it is not interested as a proof.

the 5800X3D has a lower frequency than the 5800X, but it still has more gaming performance

So, no CPU frequency? Biggest CPU cache is that I need?

[jodokus31]

Ok, I should have said, you need a CPU with high single core performance

Second time repeat: CPU frequency in Factorio is not an issue

I don't understand, what you are trying to say.

you need a CPU with high single core performance

No, I do not need it. And sorry, I did not read about your experience, it is not interested as a proof.

I didn't want to proof anything.

the 5800X3D has a lower frequency than the 5800X, but it still has more gaming performance

So, no CPU frequency? Biggest CPU cache is that I need?

CPU AND RAM (with low-latency, dual-channel and XMP) is important for factorio.
The big CPU cache helps to reduce the RAM-usage and is therefore beneficial to conserve memory bandwidth.
That's my opinion/experience, no proof.

Take you what you think is best.

It wasn't my intention to be unfriendly or whatever, but it might have sounded a bit like that. Sorry, if that's the case.

[Bloodred217]

I got a 7950X3D a few weeks ago and have been tweaking things while benchmarking Factorio. I've been keeping track of the results for myself, but now I've found this thread and perhaps somebody else would be interested to see them. I've been using my own map for testing, so these numbers are only relative. It's quite CPU heavy as you can see from the UPS not being very high, the base is "only" doing 4050SPM but it's using almost 800 small trains and 2500 stations over a large rail network which frankly makes the train pathfinder cry out in pain to the point where it often takes longer than the whole entity update.

Factorio minitrain megabase bench.png (75.84 KiB) Viewed 4322 times

Some conclusions I have reached after testing Factorio on this CPU:

• With the best settings I've found, the 7950X3D system is about 84% faster than my old 5950X system on this particular map
• Outright disabling the 2nd CCD as I've seen in some reviews does not seem necessary
• Tweaks can have a pretty significant impact on performance over a default system, from 101-102 UPS to 114 UPS, scaling is surprisingly good
• Factorio must run on the V-cache CCD, as expected
• The AMD scheduling driver works correctly if and only if the game is actually rendering graphics, as in if you're actually playing the game with a graphical window being displayed. When you're running Factorio as a benchmark in the background, it is not scheduled correctly! Still this is somewhat good news, when you're playing you will get good performance automatically, but it means many benchmarks you find in reviews effectively show unrepresentative results.
• Setting cache priority from the BIOS performs slightly better than setting CPU affinity from Windows, I don't exactly understand why. Still this is good as this is a global system setting and would apply to Linux or any other OS as well, so you don't need Windows, or a special driver, or to manage affinity yourself to get good results

Note: the faded results are slightly lower likely because I didn't close all other programs when I ran those tests

[azesmbog]

I've been keeping track of the results for myself, but now I've found this thread and perhaps somebody else would be interested to see them.

Wouldn't it be difficult to test several maps on a factoriobox?? The standard map does not interest at all, it is not interesting.
This one, although the results for 7950X3D are already available, but so far only Intel has 100 UPS
https://factoriobox.1au.us/map/info/af7 ... 9605e8d509

This map is of interest, a very small result at 12900
https://factoriobox.1au.us/map/info/91c ... 9eb9d399f9

This map is also of interest, a very small result on AMD processors.
https://factoriobox.1au.us/map/info/ecf ... 851011fc83

Well, if you have at least 64GB (preferably 128) of memory - I would like to check my map))

[Bloodred217]

Never used factoriobox before. Here are the results:

First map 96 UPS
https://factoriobox.1au.us/result/058f4 ... 6c022fb777
Second map 38 UPS
https://factoriobox.1au.us/result/6a7bb ... 4248eaf204
Third map 88 UPS
https://factoriobox.1au.us/result/24b39 ... 5211b793d3

[azesmbog]

Good results, thanks.
They confirm that the heavier the map, the smaller the discrepancy between the top Intel and AMD, and sometimes Intel is still ahead))
Well, since you have 64 GB of memory, you can try to run my map, not the most up-to-date and latest, but one of the last saves.
https://drive.google.com/file/d/1biDRq4 ... share_link
save from February 11
I'm only interested in UPS when I'm idle and doing nothing)) It will probably be enough for me .. (my UPS at the moment is ~ 32-33)

[FuryoftheStars]

Hmm, I'm noticing that while Intel is ahead, they did the tests with faster RAM (5600 vs 7800 MHz).

[Bloodred217]

Good results, thanks.
They confirm that the heavier the map, the smaller the discrepancy between the top Intel and AMD, and sometimes Intel is still ahead))

Sure, no problem! Perhaps the game is becoming more and more memory bottlenecked when the maps get larger.

I'm only interested in UPS when I'm idle and doing nothing))

64GB was barely enough to load it (took like 20 minutes with plenty of swapping) but once I was in and it settled, I got about 43-44 UPS.

Screenshot

20230415041230_1.jpg (1.28 MiB) Viewed 4209 times

Hmm, I'm noticing that while Intel is ahead, they did the tests with faster RAM (5600 vs 7800 MHz).

From what I've seen on those results on the maps posted above, AMD still seems to be slightly ahead, though the differences are so small I'd call them irrelevant anyway. Faster memory could certainly help, but note that the factoriobox script seems to get memory speed wrong sometimes, for instance mine is running at 6000 MT/s but factoriobox only reads it as 4800. I assume a lot of other memory speed entries must be wrong too.

[azesmbog]

Thank you very much for testing my map.
Approximately, I was counting on such a result If I had such an UPS now, I would still play on my map for a year or two, but since now - alas...
And yet, with 128 GB of memory in MacOS, the map loads in 2 minutes ...
Regarding some inaccuracies in displaying memory performance on Factoriobox. Yes, there is such an inaccuracy in the scripts, but in general this does not affect the final result (in the UPS), it only affects the false perception of the memory estimate These are already problems of operating systems that give incorrect results of the memory frequency.
Thanks again for testing

[Bloodred217]

Thank you very much for testing my map.

You're welcome!

And yet, with 128 GB of memory in MacOS, the map loads in 2 minutes ...

Yeah, I'm not surprised, loading the map was a real struggle. It's interesting to me that Factorio seems to need more RAM to load a map than to run it, once it settled it only used about 47GB RAM and my PC didn't need to swap, but while loading it used the whole 64GB.

Yes, there is such an inaccuracy in the scripts, but in general this does not affect the final result (in the UPS)

Yes, I'm sure it doesn't. I didn't mean to imply that the UPS displayed is inaccurate.

These are already problems of operating systems that give incorrect results of the memory frequency.

I dug around my system a bit to figure out where the 4800 MT/s was coming in my case, I think what the script is actually doing is it's getting the JEDEC profiles on the memory rather than the settings it's actually running at. Windows otherwise knows the correct speed (for instance it displays the correct number in Task Manager).

[azesmbog]

Displaying memory speed is still OS dependent. In Windows 7, the script always correctly showed the current speed (it was 3740 MHz), now in Windows 10 it shows the nominal 2400 MHz (although the real 3491 MHz is very difficult to overclock with such a volume ))
As for the volumes, it’s much more interesting, 125 GB is used immediately when loading, if you stand in one place and do nothing, it drops to 67 GB after 5 minutes, and if you still stand, it has now dropped to 31 GB.
Further. After the resumption of traffic and the start of construction, at first everything was fine, consumption almost did not grow. But when allocating a site for deletion, and then canceling this action, the game hung well, but the allocated memory began to slowly increase, in about five minutes it reached 48.3 GB and that’s it, and it doesn’t grow any more, and hangs tight (Hello developers!))
It is necessary to remove the task But it's fun to play )
Therefore, when memory consumption drops below 115-120 GB, this does not bode well for my map:)
By the way, I loaded a completely fresh save, if you're interested - you can check it) From the oddities - under MacOS it loads in 2 minutes 50 seconds, in Windows 10 it loaded in 1 minute 40 seconds)
(the application continues to hang, it is necessary to kill, it is useless to wait further)
upd:
In MacOS, the same save has been running for an hour and a half. Memory consumption did not fall below 116 GB. And it didn't hang. On the same hardware. I won't draw any conclusions, I'll just keep playing. Yes, the UPS is 32.8, at least a little. but higher than under Windows.

[Bloodred217]

Very interesting that the game uses 125GB when loading if it actually has that much memory available. No wonder it took me 20 minutes to load and plenty of swap.