Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[MaximumZero]

That's pretty close to me, Lukas. I have a single core Celeron D @ 2.97Ghz & 1.5Gb ram. I'm usually sitting at about 25-35 FPS in a normal game, but couldn't pass 12 in the test.

[Caldfir]

32-45 fps

AMD Phenom II 965 @ 3.4GHz

[janglur]

if it wuld be that ram dependent, my 7cl 1600 mhz tri-chan ddr3 ram *shuld* pawn you even with lower cpu frequencys.  still i get lower fps on my i7 920 @ 3.8 ghz, forced to one core or not

Actually, yours is CL7, meaning 7 cycle latency per read/write.  Although the frequency is faster, resulting in higher overall MB/s, that is NOT advantageous to Dwarf Fortress as it uses a lot of very small writes.  It's like harddrive latency.  A 0.45ms 50 MB/s SSD will outperform an 8.5ms 120 MB/s harddrive due to the seek time difference, as the workload is primarily seeks.  It's one of the unfortunate tradeoffs for computer performance these days.  It's what supposedly made RDRAM so superior (however it was 16-bit memory compared to the 128-bit DDR standard, thus resulting in much less throughput despite a higher frequency at the time).  That makes me want to test this old Rambus P4 of mine to see if it makes a big difference.  I doubt it though due to the sheer CPU bottleneck.


Also, Caldfir:
What OS are you using, and have you aggressively disabled services/apps?  I disabled every non-essential service and my antivirus for the test.  I'm sure you can meet or beat my score if you do the same, and are running XP.  Win7 should also perform well, but Vista is a real hog.  Microsoft underestimated the tech barrier when they released Vista.  I think they expected PCs to come a longer way in performance than they have.  I did read somewhere that 2010 may be the first year Gordon's Law doesn't get respected.  (I think it was Gordon.. the one that every year, computers become twice as powerful?)

Overall, i've noticed that sheer throughput and cache make the biggest difference.  I'm willing to wager that server-class CPUs with 8+ MB of L3 would likely perform very well.  Or a ton of L2 and L1 per-core, even better, but less existant.

[devek]

Moore, lol.

[G-Flex]

What OS are you using, and have you aggressively disabled services/apps?  I disabled every non-essential service and my antivirus for the test.

How much has disabling unessential services ever mattered? Seriously, I can't see that making very much of a difference at all. I never see crap like that really ever using a significant amount of CPU time.

[janglur]

It has less to do with CPU cycle time useage and more to do with L1/L2/L3 cache useage.

It's like RAM:  How much does it matter if you have super-speed RAM, if you don't have enough?

[Thief^]

Disabling services is all bullcrap janglur...
CPU/cache etc isn't reserved for every process/service etc, it's pretty much just used by the currently active thread.
The majority of a pc's services/processes/threads are actually asleep, not using any cpu or cache. Hell, Windows is pretty aggressive about swapping them out to the pagefile, so they're often not using RAM either. e.g. Antivirus software will be asleep until a file access happens, and DF doesn't do file accesses while running (except autosave, which doesn't really count), so disabling antivirus should make zero difference to the speed of DF.

I strongly suspect you're right about the performance difference on your machine being due to tight ram timings and large cpu cache, and DF doing a lot of random-access stuff. I suspect the way data is organised in memory will cause cache thrashing when pathfinding in some directions, most likely up/down and north/south.

Good cache performance really is a black art.

[janglur]

You're only partly right.  I can confirm a regular 10% FPS improvement by dedicated DF to a core and undedicated everything else.  They may be idle but the moment they're not, CPU and cache ARE used.  And when you have such a limited L1 as 128K and L2 as 512K, every bit helps.  (The L3 is 6 MB, but that's shared between cores.)  And remember that L1 is typically 5x times faster than L2, and L2 is 2-5x faster than L3:  L1 is usually 50-150x faster than L3.  Thats also assuming a perfect cache-hit rate, a perfect associativity scheme.

If the system was Single-Core, then disabling services would NOT help, since no matter what there's the same cache.  However, since DF is largely single-threaded, giving it a dedicated CPU core is the best possible way to max performance on a multicore system.  This prevents anything, ever, from using that core or cache for anything except hardware requests and other sub-kernal tasks.  Even system processes, services, explorer, etc. can be restricted using Process Explorer.  It's a very handy tool.

Plus, I can overclock that core to as high as 3.6 GHz stably, and the others at 800 MHz.  (However synchronously it maxes at 3.3 GHz.  3.4 works but gives minor errors after a day or two in prime benching.)

[devek]

Well, that can backfire in a way.

For example, the i7 the turbo boost stuff and the different cores have frequency bins. On my laptop core 1 has 9 bins, core 2 has 6, core 3 and 4 have 1 bin.

Putting DF on the second core might take away some of the advantages of turbo boost.

Something to think about.

[janglur]

I'm really not familiar with how Turbo Boost works, and the jury's still out on just how effective it is in a real environment.  Most likely it will be useful only in specific instances, like Execution Trace Cache and Hyperthreading.

But for an AMD system and non-TurboBoost intel, where the cores are all equal in capability, the aforementioned trick works wonders to squeeze an extra 10% out of DF

[motorbitch]

[sigh] sighs [/sigh]

[mipe9]

I don't know how good it is, but after running for a bit over a minute, it was at somewhat consistent 42 FPS (every now and then it would go to 30-ish FPS and then back to 42 FPS after few seconds)

AMD Phenom II X2 555 Processor ~ 3.2GHz
4 GB of RAM.

[jfs]

I think it's pretty clear that not just CPU type and its speed matters, but also amount of RAM (there has to be enough of it) and the speed of the RAM, as well as graphics settings and OS.

I'd suggest making a clear guide to getting the precise data using a standard method, such as reading off values from CPU-Z so everyone can post the same, making a savegame where amount of RAM matters less (small world, small embark, maybe some cave levels disabled) and using 2D or 2DASYNC mode to (mostly) take graphics card out of the equation.
Even better, maybe supply several init files with different settings, to compare the effects of them on different systems.
The current data people are posting are just too incomplete for practical comparison.

[Caldfir]

Hey, I'm bumping this monster back up to see if anyone with a new sandy bridge intel can give some numbers, as I am interested to know. 

As has been mentioned, this test is not perfect, but it is at least standardized. 

[MaximumZero]

That's pretty close to me, Lukas. I have a single core Celeron D @ 2.97Ghz & 1.5Gb ram. I'm usually sitting at about 25-35 FPS in a normal game, but couldn't pass 12 in the test.

Current Rig: 37-48 FPS.

AMD Phenom II X6 1045T @2.7GHz, 8GB RAM, Raedon HD 6450

Added model number and video card.