I noticed that Wendel makes a comment about playing Dwarf Fortress at “ridiculously high single core clock speeds” in this video
However, it’s my understanding that Dwarf Fortress performance is in large part constrained by cache misses.
Here’s a post from Putnam (now one of the developers) discussing this, and they’ve discussed this problem many times on the Bay12 forums (although not since they were hired as far as I can tell)
Well, AMD has blessed us with Epyc server hardware that has >4x that amount of CPU cache in the form of the 9684X
There’s an old(ish) benchmark built for Dwarf Fortress here, although I cannot verify it’s usefulness or quality. Here’s a blog post from the creator running it in 2021, but they don’t test cache size/speed at all.
It’s substantially out of my budget range, but perhaps y’all with your excellent access to niche server hardware could test Dwarf Fortress performance on that processor against a (for example) 9700x, fulfill the 10 year old prophecy, and perhaps finally find a cure for FPS death?
The problem with this CPU is the cache is split into 12 pieces, and the performance penalty for communicating between pieces is extreme, to the point where the other L3 cache chiplets might as well be main memory.
I’m making the assumption the performance we’re after here is single threaded, I haven’t kept up on the multithreaded experiment they were running.
This is mostly speculation but I think an arrow lake with a tuned fabric and memory OC would probably be the fastest for this workload.
Surely infinity fabric + memory latency is still significantly worse than infinity fabric + cache latency? Granted, it still might be an order of magnitude slower, but it should still be substantially faster right?
It’s multithreaded now, but the bulk of the processing (like 70-80%) is still done on one core IIRC. At least I think it was that way the last time I looked into it (about a year ago maybe?)
Does Intel have an equivalent to the infinity fabric now as well? I know they’ve started splitting up their dies the same way AMD has been.
Arrow Lake are Intel’s consumer products, which appear to have less cache/core than Zen5, do you mean Granite Rapids? Intel parts might have more L2, which could matter.
Is there a good way to find out how much cache a given cpu has per CCD? The data pages on the Intel/AMD websites don’t seem to display that info.
I would agree with this on the same processor that had relatively high latencies for all operations to begin with like the non-AM5 EPYCs.
What I’m thinking is that a ring fabric and memory tuned Arrow Lake’s latency to DRAM memory would be lower than an EPYC’s infinity fabric to separate CCD cache latency would be… this would be dependent on the workload not saturating the Intel’s bus with ~>100GB/s bandwidth, which is probably a safe assumption since it sounds like it’s mostly cache misses sapping performance.
They have different kinds of ring and mesh buses among their processors, but nothing that they’ve branded like AMD has.
I’m getting off topic here a bit but I saw that in the “rich IO” variants of Granite Rapids that Intel is repurposing UPI links into PCIe lanes to get to 136 PCIe lanes per CPU which does sound vaguely infinity fabric-esque.
The Arrow Lakes do have less cache than X3D variants of Zen5 (the 285K would actually have more useful cache than say a 9950X since half of the 9950X’s cache is on a CCD that has a latency penalty (although non-EPYC AMD CPUs can have their infinity fabric tuned to help reduce latency)), but I think it’d be pretty easy to make a 265K access DRAM memory faster than a 9684X could access L3 on a separate CCD.
TPU can be useful for this, but you have to have an idea of the processor topology to really get useful information out of it:
EDIT:
Now that I’m thinking about it more, perhaps one of the newest biggest Thread Rippers would actually be best, they can have their fabric tuned to cut down on inter-CCD latency unlike EPYC, and have a large amount of L3 at the same time.
If someone has access to a 7950X/9950X, their performance counters, could run current versions of DF bound to either chiplet, and could either generate a Large world or run a shared LP save with a stable late-game fortress without pause for one ingame year, it might at least show more about the possible gains even more lower-latency cache might do.