Definition of CPU bottleneck?

Hi,

Recently I've been having some blues with my sig laptop, namely I think the low CPU core clocks are playing a factor in game framerate performances.

Here's what I really can't understand:


IF when a game is running, and it utilizes all 4 cores (albeit 1 of them is used more than the other three), and none of the cores are maxing out at 100%, can I assume that the CPU has ZERO negative effect on the game framerates and the GPU is 100% the bottleneck in that regard?

Because people always say, oh XXX CPU is the best for gaming. Do CPU differences (different clocks / number of cores) even affect game performance if none of them are getting 100% core usage on any of the cores?



Thanks.

 

Most games are single-threaded, unfortunately, That means it'll only really be able to use one core, and thus faster clocks/less cores will perform better than slower clocks/more cores.

In my opinion and knowledge, yes you would be safe in saying that the CPU might not be the bottleneck here. I know that Nvidia GPUs have a "GPU Load" sensor in GPU-Z - see if that's the case for ATI, and how much is used during gaming.

 

If it's single-threaded, why do four cores all go under load (over 50%) when the game runs though?

What I mean is, would clocks matter if none of the cores are running over 80%? Doesn't that mean the CPU doesn't need to work at its full clock capacity to not be the bottleneck?

 

How high does the usage of your 1 most heavily used core go? If it is 95%, then you are probably CPU-bottlenecked even though it doesn't actually hit 100% (it's probably playing around with the thermal management and this is what keeps it from getting 100%). If it's more like 80%, then the CPU is probably not bottlenecking you.

Usually not, but there is not a lot of research on the subject because as a general rule, only the most pitiful CPUs out there are bottlenecks to most games (some notable games like Starcraft II are exceptions). Just out of curiosity, which game is this?

That means the game is multi-threaded, although it might not be perfectly so which would cause one core to need more power than others.

If none of the cores run over 80%, you are not CPU-bound. The only possible scenario where this is not true is if the 80% is the maximum the CPU is capable of over the long term (i.e. it gets too hot and runs at lower clock speed so your sensors tell you that it is only at 80%, but in practice it can go no higher and is thus the bottleneck).

 

In some games (like Assassin's Creed), core 1 runs at ~70%, the other 3 run at ~50%. In Crysis it's even more even. However in Need for Speed, core 1 seems to hover very close to max (but never getting to it), I'd say somewhere in the 80% ballpark, the other three cores more like 40~50%.

It's Need for Speed - Hot Pursuit. With i5 460M (faster per clock) + NV GT420M (slower than ATI 5650), I get on average 1.5~2 times the framerates at max settings.

The AMD processor runs amazingly cool in games (never goes past 70 C even after long gaming session), so I don't think that's that. However TehSuigi suggests possible "hidden" CPU throttling, so I might look into that, although I'm not too hopeful.

 

This game seems to be a primo example of Amdahl's law in action.

How is it in the other games? Need for speed might just be one of those quirky games that runs much faster on Intel processors, Nvidia graphics cards, and is only nominally mutlithreaded. After a quick search I can't find any detailed benchmarks for it though.

 

It seems to run well in Crysis, which utilizes the 4 cores quite evenly, better than i5/GT420M. But it gets these stutters now and then which reduce the fps down to 15s then climb back to 20s and 30s.

In assassin's creed, the fps is on average higher, but drops much lower than my previous laptop whenever I'm looking at many moving NPCs.

 

I'd say as long as you have a modern day GPU/CPU (i5/i7 or AMD equivalent with 4xxx/5xxx and nvidia equivalent) the only bottlenecks are the developer's code.

 

Any CPU no matter how fast can't do a div/0 or escape an infinite loop lol.

while n>0 do {n=1}

 

Bottleneck:

The factor where changing its value causes an similar change to your performance.

Eg 10% CPU OC = 10% rise in performance.

The closer to the linear scaling above the larger the bottleneck.

 

So then a 10% underclock should lower the performance by ~10%.

Might be worth testing that.

You can lower the frequency of your N930 right?

 

Yeah I think there's an option in AMD OverDrive to lower the CPU multiplier. Interesting.

 

I just did it, used FRAPS 60 second fps capture, benchmarked for CPU @ 2GHz vs @ 1.6 GHz (20% difference).

The Min/max/avg for 2GHz
21/50/30.8

The min/max/avg for 1.6GHz
14/35/20.5

This is some serious bottlenecking O.O The reduction of performance is even more than the reduction in clocks!!!!

 

Have you seen if there's a CPU overhead that FRAPS is adding to your gaming? I don't use it personally, opting to use TF2's internal cl_showfps control variable.

 

Ouch! Yes, that definitely makes it look like you are CPU-bound, but then I don't understand why it is only reporting 80% usage. Does it still report 80% when you force it down to 1.6GHz?

Regardless, the moral of the story is: don't buy quad core CPUs without Turbo Boost clocked at 2.0GHz, particularly if they are based on a crippled version of a 2-3 year old architecture.

 

Yeah it still uses ~80% core 1, and 50% of the other 3...

 

Good! This means that your CPU is definitely bottlenecking you and on top of that, either whatever you are measuring usage with is incorrect or the CPU is silently throttling itself.

It's remarkable just how awful AMD's mobile quad-cores are -- they are beaten by the better Arrandale dual-cores even in fully multi-threaded workloads, they have inferior battery life and they are utterly crushed in tasks that place a heavy load on 1 or 2 threads, not just by Arrandale, but even by 2 year old Core2Duos. Their desktop CPUs really aren't so bad (they may not be the best, but they're competitive at the prices they're being sold at), but what they put in laptops is just pitiful.

 

FRAPS uses lots of system resources. If you really want to find out the true framerates it is best not to use it if there is any other way.

 

For the CPU core usages I have both AMD OverDrive, and Windows task manager, open on background, run the game for a while, then alt-tab out to look at the core usage trends.

As for the throttling in question, it's been tossed back and forth and I still don't know if it's definitely happening, or not.

So there must be throttling if none of the cores are floored (usage stuck to the ceiling) yet evidence shows CPU bottlenecking?






Also I just tested 1.8 GHz CPU clock in the game, Fraps shows 26.5 average fps, which is consistent with the previous tests at 1.6 and 2 GHz.

 

Is there an upgrade option for that socket?

Maybe you can get a new CPU from ebay.

 

Here's a list of the CPU's that *should* be compatible:

http://www.amd.com/us/products/note...ainstream/Pages/2010-mainstream-platform.aspx

I would look into a higher clocked dual core.

You can get the N620 @ 2.8GHz for $180 and sell your old quad for $210.

 

That's an interesting idea, however while some of the game are bottlenecked by my CPU because they are single/2 threaded, the majority of new games are still multithreading, and would lose performance with the N620 because overall clock would be lower.

It seems the only perfect solution is quad/dual core with turbo (cough*intel*cough). Well that won't happen soon unfortunately, and I'm stuck with this laptop haha.

I wonder if unlocking the HT reference clock (the AMD equivalent of FSB) is possible, because that's a direct route (and pretty much the only way) to overclocking this CPU. My theory is, the guy who claimed to have overclocked this CPU to 2.5 on newegg, must have found some form of shady software which unlocked the HT ref clock and allowed him to go through with the deed.

 

Jerg, have you tried k10stat? Not had anything to do with AMD OC so apologies if it's a bad suggestion.

 

Indeed. If your measurements of what the other cores are doing are accurate (i.e. the load on them is around 50%), I doubt a faster dual-core will help you here. This game is already decently threaded and AMD's CPUs lack hyperthreading so if you try to dump everything on those 3 cores onto a single one, you'll probably be no better off than before (maybe even worse). The bottom line is that you need a modern mid-range mobile CPU and AMD has only made low-end ones for the past 4-5 years.

 

I tried it last night, it almost seemed as if I could have bumped up above 2.0 GHz, but a quick check with Intelburn + cpuz confirmed otherwise. The CPU voltage is capped already (returns back to orig clock), and if I push the clock above 2 GHz, the max clock the CPU does is the step below (1.7 GHz).

It seems the reason why none of these work is that several things are all locked already (CPU voltage, HT ref clock, CPU multiplier).

Do people who use setFSB, with similarly locked CPUs, actually unlock their FSB/HT ref clock? I know setFSB won't work here but if locked FSB can be unlocked via software, then HT ref clock should too, if a software is ever made to do that.

 

Hyperthreading wouldn't do anything in this case. It is poorly suited to increase performance for things like games. The real problem is that this single game poorly takes advantage of multiple cores. Switching to a considerably higher clocked dual core will increase performance in this game, but will be detrimental to other games and programs that are better able to take advantage of the added processing power of more cores. Intel works around this conundrum with turbo boost. Unfortunately AMD does not have any analog to it. If the AMD processor could increase the frequency when utilizing less than all four cores, the OP's issues would be resolved. This has nothing to do with your belief that AMD only makes "low-end" processors.

 

Its impact varies from program to program, but I suspect in this case it would help a bit. In any case, it is not correct to generalize to things like games.

Then why is the CPU load on the other 3 cores 50%? The number of programs that can be perfectly multi-threaded even theoretically is small (see the Amdahl's law link in this thread). It is far more common to have one or two main threads and offload the rest to others which is what the game is doing. Switch to a dual core and the load from those 3 cores has to go somewhere.

They have a rudimentary version on the hexacore desktop Thubans and I suspect they'll put it into Llano as well, but yes, lack of Turbo Boost is one of the big problems with their current laptop lineup (though by no means the only one).

No, they would not be resolved. For the last time, it is already using all 4 cores with at least half the maximum load on each. Turbo Boost would not help at all here. And I did not say that AMD only makes low-end processors, I said it only makes low-end mobile CPUs. Their desktop chips are well into the mid-range with the Thubans possibly even high-end.

 

Sigh, I just found out (according to the 5553G schematic here) that my mobo is hardwired NOT to allow overclocking - it doesn't even have a clock generator!

All the clocks are generated by southbridge instead, how "ingenious" of AMD.

Can anyone here confirm that this is indeed the case?

 

No, it wouldn't. Games are the applications least able to take advantage of it. The ability of something to benefit from hyperthreading is directly related to how parallelizable it is. Straight forward number crunching with no surprises sees the most benefit, with as much as a 30% increase in performance. Something that is already bottlenecked by a single thread, especially on a quad core processor, is going to see no to negative improvement with hyperthreading.

His processor theoretically has enough muscle to do much better in this game and this is the only game where he is experiencing this problem. Just because a program can spread a little bit of its excrement around a lot of processors does not mean it does a good job of utilizing more than 2 cores. A higher clocked dual core would continue to put the one thread that is bounding performance on a single core, and throw the rest of the work onto the second core. Performance would improve by a good fraction of the relative speed increase of the processor.

And dude, I posted the link the Amdahl's Law. And like I said before, this game is a perfect example of it in action. It doesn't matter how many extra processors you add, regardless of the extra load being divvied up equally between them, if performance is bottlenecked by a single thread.

If one thread on one core is bottlenecking the whole thing, it definitely would. Intelligently loading the cpu could throw the program onto two faster cores which would speed it up considerably.

 

Here are what happened later:

I was able to return the laptop free of restocking fee (minus shipping), and purchased another Acer, the 5820TG, which had the same specs (the laptops even look the same) except an Intel i5-460M CPU instead of the AMD N930 CPU; it costed the same because of a price-match.

Got the laptop earlier today, tested on the same games with same 5650 overclock clock values. Wow, CPU bottlenecks of the previous laptop cannot be more clear.

On average, my framerates doubled across all of ~8 games I've tested so far. One or two games which were entirely CPU-independent (e.g. Two Worlds II) I got the same framerates (meaning the N930 did not bottleneck), however in games like Need For Speed, and Assassin's Creed, I got double the framerates (40s and 50s instead of teens to low 20s); and in some extreme cases - such as Half Life 2 Cinematic Mod - I got about FIVE TIMES the framerate boost, going from low 20s to 30s FPS with the returned laptop, to now 80~200 fps.

All of these because of a different CPU.




Lesson learned (hopefully those who want to buy laptops with AMD CPUs for gaming in the future can appreciate the pain I've been through too), AMD is a no-no in laptops; nerfs it to hell doesn't matter what the other specs are.

 

you only have two cores, each with 2 hyperthreads. thus, two of those hyperthreads each at 50% == one core at 100%. as all 4 hyperthreads show 50% that means all two cores are at 100%.

and besides, even if only one core is at 100% (shown as 50%), and the others aren't, that can mean that the game bottlenecks, as at this point only one thread of the app would need more juice than the cpu delivers.

happily, to fight this single-thread-bottleneck, newer cpu's like your i5 do overclock if only one core (or simply not all cores) are at full usage, to fight against that bottleneck.


but in your case, the fact that you have hyperthreading somehow hides the fact that you have all cores at 100%.


oh, and, no, games nowadays are heavily multithreaded and use all cores. that's a myth that isn't true anymore, that games are still singlethreaded. wasn't true for halflife2, wasn't true for crysis/farcry, and isn't for a lot of other games.

 

No no no, that was when I had the Acer AS5553G with the AMD N930 CPU, which had 4 physical cores (and 4 threads), no hyperthreading, no L3 cache, no turboboost. Hence it got bottled madly.

Now I've just had this new AS5820TG for half a day yet, haven't monitored the CPU load during gaming but I'm fairly sure it's handling most games many MANY times better than the N930 processor.

 

oh, well.. points still stand. but obviously the hyperthreading part would not fit to the amd