Cache overclock increasing cpu utilisation in games

[rfcokilljoy]

Hey, I was messing around with cache frequencys recently on my i5, and found that setting the cache frequency to Auto actually provided noticable gains in Battlefield 1. 

 

With the cache at 4.7ghz (fully stable), the cpu would sit at 100% probably 90% of the time in battlefield, but on Auto (3.9ghz according to cpu Z) the cpu will sit at 100% maybe 15% of the time in game. I've tried it at inbetween values as well, and in all but the Auto setting there is clear stuttering and lower fps due to a cpu bottleneck.

 

Anyone know why? Something to do with the cpu being more balanced at 3.9? My OC is 5.2ghz btw.

[Jurrunio]

It's just like overclocking memory. By overclocking cache, you shorten the time the CPU spends waiting for cache to give the info it needs to work on something.

[rfcokilljoy]

1 minute ago, Jurrunio said:

It's just like overclocking memory. By overclocking cache, you shorten the time the CPU spends waiting for cache to give the info it needs to work on something.

I know that's what it does. But it goes against common sense for it to be faster at a lower speed. Anyone got an explanation?

[agent_x007]

12 minutes ago, rfcokilljoy said:

I know that's what it does. But it goes against common sense for it to be faster at a lower speed. Anyone got an explanation?

It's probably because of syncing issues.
Since cache works at different frequency, you have to somehow synchronise it to both slower (at idle), and faster (at load), Core Clock.
When sync can't be done, or it's timing is off, you get problems.

Basicly :
5,2/3,9 = 1.3(3) (so you get one and 1/3 of CPU clock per Cache clock OR 4 CPU cycles per 3 cache ones)
5,2/4,7 = 1,106382978723404... (you get 1 and 1/10 of CPU clock + waiting a tiniest ammount to get a full sync).

At least that's what I think.
I can't test it because Sandy/Ivy class CPUs have UnCore synchronised with Core Clock.

[rfcokilljoy]

7 minutes ago, agent_x007 said:

It's probably because of syncing issues.
Since cache works at different frequency, you have to somehow synchronise it to both slower (at idle), and faster (at load), Core Clock.
When sync can't be done, or it's timing is off, you get problems.

Basicly :
5,2/3,9 = 1.3(3) (so you get one and 1/3 of CPU clock per Cache clock OR 4 CPU cycles per 3 cache ones)
5,2/4,7 = 1,106382978723404... (you get 1 and 1/10 of CPU clock + waiting a tiniest ammount to get a full sync).

At least that's what I think.
I can't test it because Sandy/Ivy class CPUs have UnCore synchronised with Core Clock.

That is actually a brilliant answer, and would make sense. If anyone can confirm this that would be great!

[Jurrunio]

17 minutes ago, rfcokilljoy said:

But it goes against common sense for it to be faster at a lower speed.

You said it performs better at 4.7GHz than 3.9GHz didnt you?

[rfcokilljoy]

2 minutes ago, Jurrunio said:

You said it performs better at 4.7GHz than 3.9GHz didnt you?

Maybe I put it poorly, I meant at 3.9 it is at 100% usage less of the time.

[Jurrunio]

3 minutes ago, rfcokilljoy said:

Maybe I put it poorly, I meant at 3.9 it is at 100% usage less of the time.

So 3.9GHz is worse than 4.7GHz right?

[agent_x007]

1 hour ago, rfcokilljoy said:

I've tried it at inbetween values as well, and in all but the Auto setting there is clear stuttering and lower fps due to a cpu bottleneck.

@Jurrunio ^This is the important part.
Any setting other than Auto = stutters and has lower FPS (?).