Print

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

[wierd]

A thought occurs to me.

In the DF Fortress mode forum, a user named Larix has been working on and sharing plans for a general purpose dwarfputer. (now under 7000 mechanisms!) It's a bit slow, but it could be used as a benchmarking tool.

When active, it takes a healthy framerate of ~20fps down to abysmal 4-5fps, which means it is really putting DF to the screws to simulate it.  Its computation speed is pretty constant (measured in game ticks), so if we run a computationally intensive program on the dwarfputer, and time how long it takes in actual seconds to complete, we can get a rating on a host computer with DF under heavy load that is consistent.

He provided a savegame with a completed instance of this dwarfputer in the most recent thread he created. (He's been refining the computer for over 6 months now, implementing leaner/more aggressive performance related improvements over that time. It's a fully programmable Harvard architecture computer capable of general purpose computing with a reduced instruction set. He's defined the instruction set and how the program tape works in the thread.) We could task it with say, the "queens" problem, and see how long it takes in realworld seconds, and use that as a metric for CPU-generation performance when playing DF.

Just an idea.

[superbob]

That sure sounds dwarfy, but it would primarily test the performance of minecarts and mechanisms, which would not appear under ordinary circumstances in such quantities. AFAIK the most common causes of FPS death are pathing and clutter, which is why my initial ideas revolved around stressing these parts of the game.

Obviously it's possible to bog down the game in non-standard ways, but that's beyond what we could hope to cover in a reasonable amount of benchmarks. Rather, the idea is to simulate a reasonably badly designed fort, and/or a reasonably cluttered one, while using a reasonably simple macro/blueprint to generate the test.

For example a rough equivalent of a 5 year old fort with 120 dwarves and relatively light clutter/pathing costs. Meaning if you get 200FPS in the benchmark then you should be fine going for a larger embark size or being sloppy with fort layout. OTOH doing less than 100FPS would suggest a need for small embark size and optimized layout. Obviously building a dwarfputer or lots of waterfalls and magma would make things worse to some vaguely predictable degree.

While not perfect, this might be helpful to people wondering if their computer will run DF. Or people wondering if 3x expensive, faster RAM is worth the improvement in their case.

[Kamamura]

As of today, Intel CPUs are much better. What's more there to say?

[vomov]

As of today, Intel CPUs are much better.

A simple benchmark (Matlab, data processing, four different steps in processing), between two PC's:

• AMD A8-6500 @ 3.5GHz (no OC): step 1 lasts 1 hour, step 2 four, step 3 ten minutes, step 4 half an hour
• Intel Core i7 4820K @ 3.7 GHz (new PC, no OC): step 1 lasts 3 hours, step 2 nine, step 3 and 4 combined last less than five minutes

AMD seems to win in total time (4.5 vs 12), but this is is not a good measure of performance here. If I combine this with the price of the processor (€95 vs €330), or the power usage (65W vs 130W), AMD wins massively. On the other hand, a different method of processing the same data makes the AMD take around two days (!), while the Intel does it in two hours. It heavily depends on the type of processing, and DF seems to favour single thread performance, which means Intel wins.

but it would primarily test the performance of minecarts and mechanisms

Yes, but the signal (minecart and mechanism performance) to noise (everything else) ratio is quite nice here, meaning we can somewhat separate these types of load from the rest. Similarly; a fort with twenty mist generators would test the type of load liquids impose.

While not perfect, this might be helpful to people wondering if their computer will run DF. Or people wondering if 3x expensive, faster RAM is worth the improvement in their case.

That is the best we can hope for, I think; it would be waaaaay too much work to set up a load of testing fortresses, let them run on different PC's, collect some measure of performance, and do some serious data mining to figure out what influences what. What Miuramir said is the most helpful:

My recommendation for a DF system varies somewhat as we find out more info on DF and technology changes, but is usually close to:
1) Figure out the best cooling system that you are willing to put up with (in the case of desktops), listen to, and/or carry (in the case of laptops).  Laptop performance is almost entirely cooling dominated, and single-core sustained turbo ratings can be significantly different on a desktop with better cooling. 
2) Pick a motherboard and memory that offers the best GB/sec sustained transfer rate from the main memory to the CPU that you can budget, and can be adequately cooled by the solution from step 1.  You want at least 4 GB of memory, and preferably a bit more; more than 8 GB isn't going to help current-generation DF any, but might someday. 
3) Pick the best current-generation CPU for single-thread performance under sustained load, that has at least two real cores, while cooled by 1), that is compatible with the motherboard and memory from 2), and that fits your budget.  (While DF is single core, you want at least one more, so that the OS and so on doesn't cut into DF's time.) 
4) Give it a good power supply with a high efficiency rating and enough headroom to handle everything cleanly. 
5) Accessorize with everything else you need; most of the details are of low importance for DF's performance, although considerations like input devices and monitors can be important for the user experience.