All the Intel CPUs we mention are very similar in single thread performance at stock or cruise speed, all around 2000~2200 STP, which is already great for MAME.
It's the steps up that are bit of a black box, afaik we don't have large "turbo and OC" STP databases for reference.
IMHO it all comes down to what we wish to emulate, and it's good to keep in mind GM needs a little more room for frame_delay, also on the gpu side.
But really what's so exciting in MAME's library that it requires considerable processing power? I wonder sometimes if OC'ing my i5 for MAME would be worth the trouble (rather right now I need more GPU juice)
Following this I did a little monitoring and little OC on my i5-4690k.
With the default steps that CPU does 800MHz > 3.5GHz then up to 3.9GHz when the turbo kicks in. I've moved to limit to 4.2GHz (iirc it can go up to 4.6GHz if you do it properly, which takes time and I don't want to bother with that now)
NB: using W7 GM 0.203 D3D9ex
Purely on the games side no frills added, for instance it eliminated the last remaining little speed % drops I could still see with a couple Gnet and System22 games, it also made Virtua Fighter 2 full speed and Radiant Silvergun
playable (though still losing some % in the busiest moments) but not
all STV games.
A bunch of other polygonal 3D games worked full speed or close but that's about it for the drivers speed benefits. Note that out of the 4 cores the most used one differs from system-to-system. How beneficial having 4 vs. 2 can be? idk, but it wasn't obvious since as far as I've seen most of the load is always on one core, and eventually there's also a bit of load visible on a second one, never seen the four climb in unison or having close/similar loads together.
With GroovyMAME's lag reduction the gain in CPU processing power translated to higher stable frame_delay values, for instance where I could have cv1k games at 5~6 without much more than filter 1 + prescale 2, I can now play at 7~8 (-1 or -2 if using HLSL depending on game and settings: more effects = higher toll)
With systems that hated frame_delay before (like STV) the deal hasn't changed much, they can handle little bit more, like 3~4, but still collapse and crawl completely at the first big explosion or very busy screen.
My personal conclusion; unless there's a real interest for the rather few well-running polygonal 3D games, or a purpose to get the most out of Groovy's lag reduction for everything 2D (and compensate a bit for using HLSL), I can't see a considerable interest in seeking more CPU power than, say, that 3.5~3.9GHz / 2000~2200 STP average the Pentiums and i3's produce at stock speeds.
Someone who doesn't plan on using frame_delay or barely, and mostly stick to D3D9ex for his low lag fix, and also has no interest in the heavy 3D titles, he will be happy with the entry Pentium G4560 or any CPU new or used that can produce around those 3.5GHz/2000STPs.
And if not for playing like even the cv1k games, then any toaster will be fine (seriously what's a reasonable lower limit for decent MAME these days?)
For my own needs I probably won't ever need more than what this little OC brought me today.
Can I check the STP with passmark here? need to ckeck that for reference.