Tom's Hardware: CPU/GPU Bottlenecks

[BFG10K]

• Part 1: http://www.tomshardware.com/reviews/game-performance-bottleneck,2737.html

• Part 2: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738.html

Aside from known outliers like GTA4 and Dragon Age, in the vast majority of cases two cores are virtually as fast as four cores, and the GXT460 bottlenecks performance.

Therefore, in that selection of games at those settings, a dual-core + GTX480 would provide far better performance overall than a quad-core + GTX460, indicating yet again that the graphics card is the most important part of the gaming equation.

This is what I’ve been saying for quite some time, and mirrors my own findings.

 

[blastingcap]

• Part 1: http://www.tomshardware.com/reviews/game-performance-bottleneck,2737.html

• Part 2: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738.html

Aside from known outliers like GTA4 and Dragon Age, in the vast majority of cases two cores are virtually as fast as four cores, and the GXT460 bottlenecks performance.

Therefore, in that selection of games at those settings, a dual-core + GTX480 would provide far better performance overall than a quad-core + GTX460, indicating yet again that the graphics card is the most important part of the gaming equation.

This is what I’ve been saying for quite some time, and mirrors my own findings.

Yup which is why it took me forever to upgrade from an E8400 setup (I would have gone quad core but a 1055T can oc and/or turbo core and is only modestly more expensive than a 965BE).

A counterarg is that lower resolution = higher possibility of CPU bottleneck. But I think that at such lower resolutions you're probably at a decent framerate anyway, assuming a decent video card, so it doesn't matter if you're CPU is bottlenecking you at 60fps, it's still at a good framerate.

Edit1: I do wish they had more multi-monitor bottleneck articles out there, though. I discovered that my E8400@stock was bottlenecking me in TF2 for instance, even at 5040x1050. I was shocked because I thought all Eyefinity-level resolutions would mean GPU bottleneck.

Edit2: One thing that occurred to me was another article (can't find it right now) that said quadcores even if they gave equal fps, also gave smoother-FEELING play, possibly because of higher minimum fps. So minimum fps is another consideration.

 

[BFG10K]

A counterarg is that lower resolution = higher possibility of CPU bottleneck. But I think that at such lower resolutions you're probably at a decent framerate anyway, assuming a decent video card, so it doesn't matter if you're CPU is bottlenecking you at 60fps, it's still at a good framerate.

Also somebody that drops down big bucks on expensive quad/hex processors isn’t going to be playing games at 1280x1024 at no AA just to see 300 FPS vs “only” 100 FPS. These sorts of people are going to be cranking the settings as high as possible because that’s the purpose of a high-end rig.

 

[Termie]

I usually like Tom's a lot, but doing these tests almost exclusively on a GTX460 768 made no sense. They did test a crossfired 5870 setup in a couple of instances, but they should have done it every time. No reason at all to bother overclocking an i5 to 4GHz if they test most games only on the GTX460 768.

Look at the the Just Cause 2 numbers, for example: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738-2.html

Tell me that this test says anything at all about CPU bottlenecking.

Seriously...in some of their tests, the 4GHz machine underperforms the 3GHz machine: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738-9.html. Obviously, they're testing with the wrong graphics card.

 

[cusideabelincoln]

When disabling cores on their i5, don't the active cores still have access to the full 8MB of L3 cache? That would probably skew results for real-world scenario comparisons.

What I would like to see are these tests:

1. Equally clocked Core i3 and i5 (with two cores disabled). This should give us an idea how big of a role the L3 cache plays.

2. Compare equally clocked AMD processors. Try the Athlon II X2, Athlon II X3, Athlon II X4, Phenom II X2, Phenom II X3, and Phenom II X4. Disabling cores on just the X4 parts doesn't paint a real-world scenario, IMO, although I suspect if done on the AMD parts the difference wouldn't be that big. In other words I'm trying to say the cache probably has a bigger impact for the Intel chips than the AMD chips.

But I do agree for most games a good dual core is all you really need.

 

[toyota]

I knew somebody would finally post this. you cant just disable cores on a certain cpu architecture and draw a conclusion of how many cores are needed. I guess if I disable a core on my E8500 and a game is playable then a single core P4 could also provide the same performance. lol

here is a more realistic test that show different modern architectures can have an impact. notice how i5/i7 doesn't suffer in BC 2 with 2 cores disabled but the 965 X4 and especially Q6600 sure do. http://www.legionhardware.com/articles_pages/gaming_the_core_debate,1.html

and like cusideabelincoln hinted at, if you used REAL dual core cpus then results would be even worse in most cases.

 

[Concillian]

It's pretty much always been GPU >> CPU. Pretty much since the first 3DFX card.

The reasoning is simple.

Look at "BENCH" for an e5200 vs a 980X in single threaded Cinebench performance.... The 960X is like 1.6x the e5200. Not much of an advantage. Even if you were to take full advantage of 6 cores v.s 2, you're still at only a 4.8x multiplier from the bottom end to the top.

Now look at GPUs... nVidia's site claims the GTX480 is 33x more graphics power than the GTS210. Umm... 33x is quite a lot more than 4.8x, and designers often have to make games able to be semi-playable on an IGP even.

Designers have to develop games to be a LOT more scalable with GPU. The difference between a $50 CPU and a $1000 CPU is 1.6x in single thread performance? There isn't a whole lot of room to scale CPU performance. Even if you can program for 6 cores, the graphics landscape is 5-10x wider from top to bottom.

CPU will always be of minor consideration to gamers. Game designers just can't put the same kind of "candy" in for high end vs. low end for CPU. You only have 60% - 400% to play with instead of 3000%.

This is why I have an overclocked i3 instead of an i5 or i7... just don't need more for what I do. It really doesn't take a SERIOUS computer anymore to get perfectly playable performance at high detail settings unless you're running a 30" monitor or a 3 monitor configuration. A 4GHz i3 really isn't that limiting at all if the most demanding thing you do is gaming.

 

[BFG10K]

I guess if I disable a core on my E8500 and a game is playable then a single core P4 could also provide the same performance. Lol

Nobody is claiming that because a P4 has more architectural differences other than just having fewer cores. In particular, it has a massive IPC deficit compared to the Core.

What is being claimed is that a mid-range CPU + high-end GPU is better for gaming overall than a high-end CPU + mid-range GPU.

and like cusideabelincoln hinted at, if you used REAL dual core cpus then results would be even worse in most cases.

I used a real dual-core (E6850) and I saw similar results against my i5 750. I also saw virtually no performance gain from an i7 870 over an i5 750 despite having a 14% clock advantage (I’ll have those results up soon).

 

[toyota]

well I consider an i5 750 to be mid range and of course that cpu is plenty enough to handle any gpu configuration. with overclocking that i5 750 will almost match a $1000 cpu and will be plenty for years to come. IMO there is no point in trying to save 75 bucks on the cpu when building from scratch.

 

[blastingcap]

well I consider an i5 750 to be mid range and of course that cpu is plenty enough to handle any gpu configuration. with overclocking that i5 750 will almost match a $1000 cpu and will be plenty for years to come. IMO there is no point in trying to save 75 bucks on the cpu when building from scratch.

I agree that it's better to go for a strong CPU+mobo+RAM combo and worry about the GPU later. It's easier to upgrade a GPU than a CPU, that's for sure, especially with Intel's rapid-fire socket switching. But it's still wise to keep in mind that most games are GPU-limited at 19x12 resolutions and higher.

Oh and, i5 750 is at the upper limit of what I'd call midrange. It's a $200 CPU!

 

[fffblackmage]

For me, those article are too late. I already experienced the dual core versus quad core thing first-hand. My quad core upgrade wasn't much of an upgrade, or at least it wasn't mind-blowing.

 

[toyota]

For me, those article are too late. I already experienced the dual core versus quad core thing first-hand. My quad core upgrade wasn't much of an upgrade, or at least it wasn't mind-blowing.

well you have a 4890 so any fast modern dual core will feed it just fine in most cases. only a few games such as GTA 4 and some RTS games would have any noticeable improvement. if you were running a 5870 or especially a 5870 crossfire setup then the quad would be of more benefit for fully pushing that setup.

 

[crisium]

It's more important allocate your extra money toward a GPU then CPU for gaming, but you still need a solid baseline. I would have liked to see some 3-core results. Basically the games that see a performance gain from more than 2 cores max out at 3. Usually the Quad has a slight advantage of about 100MHz or so, So an Athlon II X3 at 3.1GHz performs the same as an Athlon II X4 at 3.0GHz.

For the budget AMD gaming rigs, it's best to get a triple core rather than a quad and put that extra money into the GPU. Not much of a concern for most of us, but seriously: 2 cores are good most of the time but 3 cores are good all of the time. When benchmarks are available, I have almost never seen 4 cores provide an appreciable difference over 3. So when you are recommending parts for your friends' budget build, the AMD triple core is pretty good for the money compared to the quad.

See for yourself, pay attention to GTA IV, the legendary 4-core game. Look at the Phenom II X3 vs X4.

http://www.tomshardware.com/reviews/phenom-ii-x2,2324-10.html

3-core to 4-core sees a 7.4% increase. Most other games are less.

This article also supports the 3-core factor, although the low settings are funky:

http://www.tomshardware.com/reviews/multi-core-cpu,2280-10.html

Yeah I'm just singing the praises of the triple core here. Don't mind me.

 

[dualsmp]

I'm wondering how Hyperthreading mixes into the equation? I noticed on my i3 530 when playing Dirt2 that Core 1 and 2 were hovering around 75% and one of the HT virtual cores was being utilized quite heavily around 85%, but the other HT virtual core was near idle.

 

[Soccerman06]

My first question is why they tested a 750 and not an i7 920-950, since you can accurately test true cores instead of 2 real and 2 virtual threads. Meh just me trying to get a real answer.

 

[busydude]

My first question is why they tested a 750 and not an i7 920-950, since you can accurately test true cores instead of 2 real and 2 virtual threads. Meh just me trying to get a real answer.

i5 750 is a true quad core.

 

[Lonyo]

I usually like Tom's a lot, but doing these tests almost exclusively on a GTX460 768 made no sense. They did test a crossfired 5870 setup in a couple of instances, but they should have done it every time. No reason at all to bother overclocking an i5 to 4GHz if they test most games only on the GTX460 768.

Look at the the Just Cause 2 numbers, for example: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738-2.html

Tell me that this test says anything at all about CPU bottlenecking.

Seriously...in some of their tests, the 4GHz machine underperforms the 3GHz machine: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738-9.html. Obviously, they're testing with the wrong graphics card.

They aren't testing with the wrong graphics card at the resolution they are using.
Actually their test is very sensible. Take a graphics card positioned at a certain resolution, and test it at that resolution.
http://www.techpowerup.com/img/10-09-06/27b.jpg

They paired it with a sensibly priced CPU, the sort which you might see in GTX460 system, which is the easiest way to see which component you should think of spending more on (~$200 GPU vs ~$200 CPU).

 

[Pott]

Mhmm... I don't know all THAT much on the topic but I'm glad I didn't spend 80 euros more to go with an i5-750 instead of the i3-540.

 

[nemesismk2]

Thanks BFG10K for posting those links, they made for an interesting read because I don't usually read THG very much.

 

[Voo]

Just skimmed through, but from what I see they only look at average framerates and no minimums which makes the whole review rather pointless :/

 

[Aristotelian]

Just skimmed through, but from what I see they only look at average framerates and no minimums which makes the whole review rather pointless :/

Take a look here for information on minimum framerates: http://www.tomshardware.com/reviews/game-performance-bottleneck,2738-15.html

Edit: That's Chapter 15 of Part 2 of the review. You even skimmed the table of contents, apparently.

 

[Voo]

Edit: That's Chapter 15 of Part 2 of the review. You even skimmed the table of contents, apparently.

No I only looked at part 1 and figured if they didn't have minimums there, that wouldn't change.. thanks. After all why not add the minimums to the particular game? My mistake

Interesting, thought that the differences there would be larger - GTA is a well known outlier, so best to ignore that. Though more than 3 would be nice - shouldn't they have that information anyhow if they use fraps?

 

[polarmystery]

It's pretty much always been GPU >> CPU. Pretty much since the first 3DFX card.

The reasoning is simple.

-snip-

Not quite...

 

[Idontcare]

When disabling cores on their i5, don't the active cores still have access to the full 8MB of L3 cache? That would probably skew results for real-world scenario comparisons.

That is one of the downsides to the "synthetic" core-count tests that THG does.

I love the trick they do, its crafty and downright easy, but yes unfortunately it does give you an oranges-to-apples comparison because of all the uncore-stuff that doesn't scale correctly with the core-count manipulation they employ.

What you can say with confidence is that it gives you the absolute best-case scenario for dual-core results that you can expect to see in the real-world.

Now if only they had a way to selectively soft-disable blocks of L3$...you can bet the tools exist at Intel (and AMD).

 

[ArchAngel777]

That is one of the downsides to the "synthetic" core-count tests that THG does.

I love the trick they do, its crafty and downright easy, but yes unfortunately it does give you an oranges-to-apples comparison because of all the uncore-stuff that doesn't scale correctly with the core-count manipulation they employ.

What you can say with confidence is that it gives you the absolute best-case scenario for dual-core results that you can expect to see in the real-world.

Now if only they had a way to selectively soft-disable blocks of L3$...you can bet the tools exist at Intel (and AMD).

Even so, we are talking less than ~10% performance difference. That is not enough to change the conclusion of THG's article.