Discussion Per Core vs All Core when overclocking ?

[SyCoREAPER]

Been doing some reading and it seems that synthetic benchmarks, which I can understand, but also game benchmark performance/scores drop, which I don't understand, on a per core OC.

(Speaking strictly Intel)

I'm curious why a higher 4 core overclock on say a 6 P Core results in less performance than all cores at a lower overclock.
Are games actually using all cores these days or what's the reason?

 

[Aeacus]

High frequency isn't everything. Computes that CPU does and how exactly CPU executes them, are far more complex than one might think.

I'm curious why a higher 4 core overclock on say a 6 P Core results in less performance than all cores at a lower overclock.

Not a definitive answer per se, but one possible reason is: parallelization.

The type of instruction being processed also matters. Modern game engines often take advantage of parallelization and delegate high-intensity tasks to multiple cores.

Certain computing tasks run faster when more cores can process instructions in parallel. Other tasks are better optimized for processors with high clock speeds. All PC games rely on both types of tasks to deliver a seamless gameplay experience.

Source: https://www.intel.com/content/www/us/en/gaming/resources/5-reasons-to-overclock-your-next-pc.html

Example:
Let's say you have 4 core OC, where the remaining 2 P-cores are running default clocks.
Now, if a game (application) sends it's compute requests to all 6 cores, whereby it is required that all cores cross-validate their results between each other, the total time it takes to execute it, is as fast as the slowest core is able to do it.
Thus, single or few core OC doesn't help at all. Only all core OC would benefit in this specific task execution.


If the above example is complex to understand, here's real life analogy to help to make sense of it;
Lets say 4 people need to get from point A to B. And they all have their own car;
* If two of the cars drive 120km/h and other two at 60km/h, then to travel 60km away, it takes total of 1h for all 4 persons to arrive. Sure, first two will make it in 30mins, since their speed is 120km/h but when it is important that they all must arrive (to go to e.g bowling), total travel time will still be 1h.
* But if all 4 cars would travel at 80km/h, it would take 45mins for all 4 to arrive.

 

[thestryker]

Someone who's spent more time on 12th gen+ can feel free to correct me if I'm wrong, but historically Intel per core clocks tend to have ratios set for a certain number of cores being used. You can set individual core ratio maximums, but I believe they're all still beholden to the overall setup.

Example (hypothetical 14900k):
2c set to 6ghz, 4 set to 5.8ghz, 8c set to 5.6ghz (stock turbo max) vs all core set to 5.7ghz

Any workload using over 4 cores would be faster on the latter due to higher core clocks under that circumstance, but anything that used 1-4 would be faster on the former.

I'm pretty sure this is still how it works, and is why in TPU's 14900K review the 5.5ghz all core OC never has a meaningful victory over the power limits removed (usually loses to stock too) for games.

 

[Aeacus]

Someone who's spent more time on 12th gen+ can feel free to correct me if I'm wrong, but historically Intel per core clocks tend to have ratios set for a certain number of cores being used. You can set individual core ratio maximums, but I believe they're all still beholden to the overall setup.

This is true. Has been so since Nehalem architecture (2008 and onwards).

Then again, usually when CPU OC is done, it is all core OC, negating the effects of different Turbo Boost ratios on different core amounts.
But on the flip side, the latest CPUs are so fine tuned that they have little, if any OC headroom (frequencies over max turbo ratio).

Back in the day, with older CPUs, CPU OC was worthwhile.
E.g i have i5-6600K with 3.5 GHz base and 3.9 Ghz boost. With CPU OC, i could get it 4.5 Ghz all core (increase of 600 Mhz over boost), or with delid, ~4.7 Ghz all core (800 Mhz over boost). And there have been some delidded i5-6600K CPUs, that can hold 5 Ghz all core.

Essentially from 12th gen and onwards, most chips out there can only hold all core stable 100-300 Mhz over max boost. That gain is so little, that CPU OC with current, highly efficient chips, isn't worthwhile. There won't be any meaningful performance increase.
If the headroom would be bigger, like it is with my 6th gen CPU, where on minimum, i look towards 600 Mhz increase over boost clocks (or up to 1.1 Ghz over boost, if very lucky with delidded chip), then CPU OC makes sense.

All-in-all, CPU OC is dying niche and outside of record breaking, isn't worth the effort anymore. Better to run stock clocks and let CPU to decide when to turbo up. Less energy waste and less heat production this way also.

 

[SyCoREAPER]

Hmm so it does seem from what I'm reading here that it's the workload that spells out of how many cores are needed.

What I find interesting is how many games to take advantage of more than 4 cores. Way back when games would only use two and a very select few would use four.

I have to check when I get home now if they are all 56x or if I have all but two at 56x and two at 58x on my 13700.

I'm kind of excited to get home and look with the information provided above. Thank you all.

 

[hotaru.hino]

I'm curious why a higher 4 core overclock on say a 6 P Core results in less performance than all cores at a lower overclock.
Are games actually using all cores these days or what's the reason?

When it comes to generating commands to render a frame, that part isn't really multithreaded. While the work submission to it can be multithreaded, there's still only one GPU, and the number of commands to be generated is capped by say 1-3 frames, so there's no real point in having multiple command generators.

As a result, getting higher FPS means needing higher single core performance. However, this is after the CPU can manage all of the other game logic in a timely manner. So if you want to target say 200 FPS, the game logic and render command generation need to happen in less than 5ms.

Some game benchmarks or other display stats show "CPU render time" or something like that. DOOM 2016/Eternal, Shadow of the Tomb Raider and Forza Horizon 5 off the top of my head have this. That's a useful metric since that tells you the fastest the game can run on your system.