Is there any downside to per core OC?

I just recently started using a per core overclock instead of syncing all the cores (I understand this is how most Intel CPUs work with turbo mode anyway). Are there any negative effects of doing this? Like one core slowing down another because they have to sync or something weird like that? My guess is that there isn't any significant impact unless all cores are at 100%?

Intel turbo's higher the fewer cores it's using yes, broadwell-E is the one that has a "Favorite Core" so to speak that picks the core that can clock the highest

Can't imagine there's really any downsides aside from more work going into it

how much higher are you getting? a few .1 ghz steps?

Although windows I don't think is smart enough to only use the fastest cores

I'm on haswell-E and have the first core at 4.2, incrementing down to the last core at 3.5.

I'm interested in knowing if the 4.2 is wasted for most typical workloads.

Only if you're worried about power consumption

why not just run the whole thing at 4.2ghz at all times?

If its a single threaded workload it'll definitely be beneficial. Or, if its one of those programs that uses many cores but still loads up on core one, it'll help. I know I saw somewhere that DX12 doesn't actually spread the load evenly across all cores (neither do Vulkan or DX11), but instead put a lot of load on core 1 and putting a bit less load on every subsequent core after.

1 Like

Heat.

What's your CPU cooler?

I suspected as much. Thank you.

Noctua D15 but it's in a rack case so air is front to back only. I tried a CLC but it made me nervous when it started to leak and the only thing that improved was the delta from 0-100% load. The idle temps stayed relatively high. I'd like to just keep the voltage down and really push the first few cores, but I didn't know if I was sacrificing anything by doing that.

For single-threaded code, it's probably great. I'd guess it might make it harder for multi-threaded code to stay synced up (whether that's a net loss probably depends on if the CPU has time switch contexts, do something useful, and switch back... without messing up the cache for the multi-threaded stuff). But that's really just a guess on my part.