Why would more cores be a bottleneck?

[Ninjasupahsquid]

So I just saw this video...

 

 

and as he is doing benchmarks I keep asking myself... why are more cores a bottleneck... surely more cores would make the bottleneck gone seeing as there would be nothing to limit the graphics card... please enlighten me

 

- Many Thanks,

Ninjasupahsquid

[Zwamdurkel]

a lot of games don't use all cores so in that case the single core performance is more important than the overall performance.

[Sernefarian]

For that particular one I do believe it is actually that each of the cores has a lower clock speed and probably slightly lower IPC in comparison to say an 8700k. These two things result in lower per core performance in games which tend to be optimized towards between 2(rarer now) and as many as 6-8(also fairly uncommon still) with 4 cores being the most common core utilization optimization point. For 99% of all games now in circulation most of those cores will go nearly or completely untouched by the games code.

[KingArchlich]

3 hours ago, Zwamdurkel said:

a lot of games don't use all cores so in that case the single core performance is more important than the overall performance.

3 hours ago, Sernefarian said:

For that particular one I do believe it is actually that each of the cores has a lower clock speed and probably slightly lower IPC in comparison to say an 8700k. These two things result in lower per core performance in games which tend to be optimized towards between 2(rarer now) and as many as 6-8(also fairly uncommon still) with 4 cores being the most common core utilization optimization point. For 99% of all games now in circulation most of those cores will go nearly or completely untouched by the games code.

This is why the whole Intel Vs AMD idea is so heated. Intel is more singular core focused while expanding to other core use as necessary, while AMD spreads itself a bit more evenly rotating core use.

 

 

[D13H4RD]

IPC 

 

It’s worth noting that while many games do have multi core support, it’s usually up to 4 cores, and there aren’t that many games that support hyper-threading. An increasing number of games are starting to support both hyperthreading and CPUs with 6+ cores but as of right now, up to 4 cores is what many games will support.

 

Intel has historically been very good in IPC performance but AMD’s Zen architecture has been closing the delta. Regardless, it’s the IPC that has people recommend Intel CPUs for gaming-oriented builds. When you add in streaming or the potential for tasks that require a heavy multi core workload however, that swings the ball towards AMD

[KingArchlich]

For the sake of making an analogy... then leaving before this gets intense...

Think of Intel as a raft along a river with some air buoys on the side.

AMD on the other hand is a string of innertubes. 

Well, Im out. Hopefully what others have said before and what comes after makes things clear.

[Enderman]

Single core performance is more important in some tasks than having many weaker cores.

[Ninjasupahsquid]

2 minutes ago, Enderman said:

Single core performance is more important in some tasks than having many weaker cores.

no... just no... It's better to have multi core performance... kinda boils down to this... would you rather fight 100 tiny Cores or 1 HUGE core

[Fasauceome]

1 minute ago, Ninjasupahsquid said:

no... just no... It's better to have multi core performance... kinda boils down to this... would you rather fight 100 tiny Cores or 1 HUGE core

"in some tasks," kinda sounds like he means gaming. Also, single threaded performance affects multi threaded, like if you had 20 cores at 1.4 GHz vs 6 cores at 5.5 GHz

[Enderman]

2 minutes ago, Ninjasupahsquid said:

no... just no... It's better to have multi core performance... kinda boils down to this... would you rather fight 100 tiny Cores or 1 HUGE core

If a program only uses 4 cores, a high single core performance 8 core CPU will perform better than a lower single core performance 32 core CPU.

Do you not understand that the program will only use the number of cores it's programmed to use?

Anything else is literally useless.

 

This is why most games and many other programs will perform better on an 8700k than a 28 core xeon.

The xeon has more cores but anything other than number crunching, rendering, folding, etc. will not use all the cores.

In many cases single core performance is what matters most.

 

PS- single core performance doesn't literally mean only one core.

It means a CPU which has a higher performance per core when compared to anther CPU.

Eg. an 8700k has higher single core performance than a 2700X.

[Ninjasupahsquid]

Oh shit... I think I just started a flame war

[Enderman]

5 minutes ago, Ninjasupahsquid said:

Oh shit... I think I just started a flame war

No, you just have trouble understanding that many programs have a limit on the number of cores they use.

That's ok, some day you'll get that having 4 big pipes is better than having 100 small pipes if something can only use 4 pipes.

 

[Sernefarian]

For random number usage that accurately describes the theme of this thread let's say we have two processors, and CPU1 is 4 core and CPU2 is 6 core.

 

Each of the cores of the four core is worth 1000 points.

Each of the cores of the 6 core is worth 700 points.

 

Application A uses only four cores for processing and CPU 1 scores 2500 points in it, and CPU2 scores 1750.

Application B uses 6 cores or more CPU1 scores 3000 points here yet CPU2 scores 3150 here.

 

... savvy? That is an accurate simplification of what this thread is saying.

Edited July 22, 2018 by Sernefarian
bah! grammars!

[Toysrme]

Cache accessing/routing issues
Memory location & routing issues, especially in the case of multi CPU (and multi-die Ryzen) you eventually have NUMA memory issues.

After you've gone past the peak, both of those tend to  combine to drop performance around 1-2% per core pair in the *best* case study games.

You have to understand, video games are not a type of program that lends itself to being threaded to run in parallel. SOME OF IT CAN, but not everything! It's all said & done a main thread has to be completed in order to advance. Even in the best modern games where main job threads can spawn as many worker threads as allowed. You reach a point where work simply will not longer scale and performance continues to degrade.

For doing things in parallel, you need to look at Amdahl's law. You can see in the best cases below (DEMD, ROTR, GTAV, anything CryEngine); for all the heavy threading the games/engines can do; only a portion of the work is really benefiting. After that, you're still waiting on work to finish before you can move on to frame generation.

Here was my study of it using my old 80 core dual xeon server from 2016. You can also see why nobody bothers much with Mantel, Vulkan or dx12. It takes industry leaders to eek out what for performance (at the expense of time=money and loosing image quality as dx11 still has wider support for various effects)

The best game that's been made so far:


Second best (after the first year of patching)


*Note about Ashes*, it is slightly better now because at the time they had imposed a hard worker thread limit of 16 worker threads. This now spawns to 40 threads (the windows standard). Altho I must say... Ashes is not so much a game. It's simply a synthetic benchmark that a half dozen companies poured their own R&D into. Also, Planatary Annihilation's latest patches scale even better than ashes. (Unit counts are cray-cray in PA:T)


GTAV is fantastic peaking at 12 cores considering it was designed to run on a 3 core xbox 360 or a 1 core + 7 baby core cell CPU in a ps3.


SC is a cluster-mess, but represents the same as any CryEngine game. Performance peaks at 8 cores (Yes even after 3.0)






Civ has always had the reputation of being threaded. It's not. That's a complete fallacy.

[Sernefarian]

IIRC Intel is having some difficulty migrating Ringbus to the next manufacturing node, or at least had been which has repeatedly delayed their 10nm advancement. Wondering if they squashed that with an alternate method. Watched some vids and read some articles on interesting potential solutions based on differing configurations where some of them outperformed that used for RIngbus by a fairly noticeable margin at the cost of increased manufacturing complexity and potential interposer use for some of them.

Edited July 22, 2018 by Sernefarian
Yes it is off topic-ish. my bad.

[jones177]

Not really bottlenecking.

 

With my 6 core i7 8700k average usage is 17%. With my i7 6700k & i7 2600k 4 cores average usage was 24%.

The performance gain in gaming is all IPC. 

 

Here is a good video about core usage in games. 

 

For picking a CPU for the best gaming performance I use Cinebench single core chart. https://www.cpu-monkey.com/en/cpu_benchmark-cinebench_r15_single_core-7

 

[Sernefarian]

For my massively simplified analogy using those cinebench core scores I found a 4 core 6 core processor set that pretty closely matches the score numbers I used in it. lol

 

https://www.cpu-monkey.com/en/cpu-intel_core_i7_7700k-664 for the 4 core

and

https://www.cpu-monkey.com/en/cpu-intel_core_i7_5820k-440 for the 6 core

Edited July 22, 2018 by Sernefarian
sentence edit.