Friday, October 8, 2010

The Game Rundown: Finding CPU/GPU Bottlenecks, Part 2

We're back with Part 2 of our bottleneck exploration in a mainstream gaming PC equipped with a Core i5 and GeForce GTX 460, ready to dive into test results from ten additional games. All of the results are summarized and analyzed in our conclusion.

Let’s reiterate a few background points for this two-piece article. We decided to analyze many popular PC games across a range of hardware and settings in an effort to identify key bottlenecks you might encounter on a fairly mainstream gaming PC. Do you need more than two processor cores for immersive gaming? Will a powerful graphics card work well, even if the CPU is weak? How much CPU and GPU performance do you really need?

In Part 1 of this article, we looked at Alien vs. Predator, Alpha Protocol, Anno 1404, Avatar, Battlefield: Bad Company 2, Bioshock 2, Call of Duty: Modern Warfare 2, Divinity 2: Ego Draconics, Dragon Age: Origins, and GTA IV EFLC. All were tested on an LGA 1156-based system. The Core i5-750 processor was configured with one, two, and four active cores and a 3 GHz clock speed. Additionally, we tried out four cores overclocked to 4 GHz. For graphics, we used a GeForce GTX 460 card, which provides sufficient performance for a solid gaming experience, but we also added a AMD Radeon HD 5870 here and there to look at potential benefits when pursuing high-end graphics.

So far, the results make clear that two cores are usually enough for gaming, but many games have specific requirements and performance characteristics. Let’s now look at ten more games and summarize the findings.

Nvidia and ATI Graphic Cards
CPUIntel Core i5-750 @ 4 GHz (21 x 190 MHz), Lynnfield design, 1.26875 V core voltage, 45 nm, LGA 1156
MotherboardGigabyte P55A-UD7, PCIe 2.0, 3-Way SLI
ChipsetIntel P55 Express
MemoryOCZ3G2000LV4GK, PC3-16000 Golden Series, 2 x 2 GB DDR3, 2 x 570 MHz 7-7-7-19 Timings
AudioRealtek ALC889
LAN2 x RTL8111D
HDDsSATA 3Gb/s, Western Digital Raptor WD300HLFS
DVDGigabyte GO-D1600C
Power SupplyCooler Master RS-850-EMBA 850 W
Drivers & Configuration
GraphicsATI Catalyst 10.7, Geforce 258.96
OSWindows 7 Ultimate 32-Bit
DirectX9, 10, and 11
ChipsetIntel 9.1.1

All test results are presented in two different charts. The first shows CPU and GPU utilization as percentages. The second lists GPU performance with the number of active CPU cores. The results are not sorted, and the order remains the same throughout the article. We start with Intel's Core i5 CPU with one active core (black bar), followed by the GeForce GTX 460 GPU (green bar), and work our way up to the overclocked quad-core configuration, just as before. In each case, the number of bars corresponds to CPU cores, and average utilization is shown as a percentage value.

One more word about CPU performance: the load may be divided between several cores in a number of these benchmarks, but the only relevant thing to notice is the level of utilization. Just because the load is distributed between all four cores does not mean that the game fully supports multi-core CPUs. Ideally, all cores would then be utilized at close to 100 percent. Load distribution and management eats some CPU power. You'll be able to see quite clearly when all four cores are really fully utilized in a game like Grand Theft Auto 4.

In the second chart below, you see the frame rates of our different CPU and GPU combinations. Just Cause 2 suffers from rather low frame rates in general, and by looking at the utilization numbers, it's hard to grasp exactly what might be causing the bottlenecks. The graphics card seems to be the most definitive limiting factor, reaching at least 92% utilization in all different CPU/GPU combinations.

According to the test results, a dual-core CPU should deliver better frame rates, as the single-core CPU is fully utilized in this game. But despite the dual-core CPU having a much higher utilization (more processor work getting done should lead to higher frame rates) there is no visible effect on performance numbers.

Despite running with 8xAA, just 573 MB of graphics memory is used. This doesn't seem to be a CPU-limited game, so we would recommend going with a dual-core CPU, then upgrading the graphics card for better frame rates.

Using a single-core CPU limits graphics card performance in this game, while a dual-core CPU is perfectly fine. Overclocking does not result in any frame rate increase, but CPU utilization goes down.

With anti-aliasing activated (the game only supports 4xAA), 585 MB of graphics memory gets used. You need at least two CPU cores to get the most out of a GeForce GTX 460 in this game.

In Mass Effect 2, you immediately notice the improvements made to the Unreal engine. While Alpha Protocol and BioShock 2 are content with a dual-core CPU, Mass Effect 2 needs at least four cores. Looking at the frame rate increase and hardware utilization tells us that this game would happily make use of an even faster CPU. Overclocking the processor yields somewhat diminished returns, though, with utilization dropping slightly. The game actually seems to treat our 3 GHz quad-core configuration as something of a sweet spot.

With this game, you should grab a quad-core chip and then start looking at graphics card upgrades. But don't lose any sleep over this issue. All hardware combinations tested produce smooth frame rates. We couldn't set the anti-aliasing to 8xAA via the in-game settings, so with AA deactivated, the game uses just 454 MB of graphics memory.

This game will not even start with the CPU set to employ a single core. Looking at CPU utilization, though, we see that two cores are sufficient and there's no advantage to using four cores at all. Overclocking yields some benefit, with utilization dropping and frame rates increasing ever so slightly.

Game play is pretty smooth, but to get even better frame rates you should suffice with a dual-core CPU and then focus on a better graphics card. As for graphics memory, the game only uses 600 MB, even with AA activated.

If you play StarCraft II with a single-core CPU, you won't get the most out of a GeForce GTX 460. The game is rather content with a dual-core CPU, and while it leverages a quad-core CPU, only two of the cores are fully utilized, with graphics card utilization climbing from 93 to 96 percent. Overclocking yields a slight decrease in frame rates. This brings CPU utilization down somewhat, while graphics card utilization goes up to 97 percent.

We're a bit low on graphics card memory reserves with the 768 MB GeForce GTX 460, as StarCraft II guzzles 611 MB, despite the lack of AA. To get more out of this game, you should upgrade to a dual-core CPU and then focus on the graphics card. Fortunately, an upgrade doesn't make much sense for StarCraft II. Frame rates are smooth with a dual-core CPU and an Nvidia GTX 460, even in single-player or replay mode.

When using the default settings without any sort of .ini file tweaks, Wolfenstein doesn't require much hardware horsepower. For this test, we disabled the frame rate limit, as we'd otherwise flatline at 60 FPS. With two cores, the game feels really natural, and the graphics card is fully utilized. Judging by frame rate values, you can even manage with a single-core CPU.

If you have a weak graphics card, you should upgrade it and then go for a dual-core CPU. Graphics memory utilization is minimal, at just 379 MB due to the lack of AA.

Since our 20 games are very different when it comes to CPU/GPU utilization and dependency, we have to be very cautious when drawing any conclusions. No single answer can be applied to all games.

Take Alien vs. Predator as an example. You can easily play it with a weak CPU and a strong graphics card. But using Grand Theft Auto 4 EFLC with a setup like that would severely limit frame rates, since it's so CPU-heavy. And while budget hardware can hit more than 60 FPS in old games without breaking a sweat, the DirectX 11-class Metro 2033 and S.T.A.L.K.E.R. titles are unplayable without a good graphics card.

There are a few things you can take away from these graphs. Even if a single-core CPU is only 88% utilized, you are still missing out on about 30% of your potential graphics card performance. With a dual-core CPU, the untapped potential is just 9%, and a quad-core CPU can bring this number down to around 5%. It's important to note that multi-core CPUs decrease the jerky gameplay you might suffer on a single-core system when games need to load new resources. In any case, all four cores are very rarely fully used, leaving software with some breathing room.

In this overview, all results from the different benchmarks are added up without weighting. All values are in frames per second. We decided to not convert them to percentages, since the results tend to point towards dual-core CPUs, although most new purchases will likely be CPUs with at least three or four cores. As you can see, the performance gains from overclocking the CPU are almost non-existent at these high graphic quality settings, the 1920x1200 resolution, and with a GeForce GTX 460.

In practice, Supreme Commander 2 cannot use all CPU cores. So one interesting question becomes whether we can we can reduce power consumption by disabling some CPU cores in the BIOS. The first chart below shows the different stages of disabling cores. On average, you can save up to 28 W overall by slashing the Core i5 down to a single core. Of course, we don’t consider this a worthwhile and realistic scenario, but the results are interesting from an analytical standpoint.

In the second chart, we disabled CPU cores in the game Anno 1404. The difference between two and four cores is the only noteworthy one, since the frame rate drop is rather large when going single-core. Thus, the power savings isn't worthwhile. If you don't want to play with reduced frame rates, you should stick to two cores, saving around 13 W on total system power.

All measurements are taken at the power plug. The Cooler Master RS-850-EMBA power supply we used features about 82% efficiency.

Why you should switch to a powerful multi-core CPU? Loading times could be one reason. Differences can be quite significant, and the more complex a game is, the more impact the CPU has. The shorter the bar in our graphs, the faster the game loaded. All numbers are in seconds.

Overclocking the CPU finally has an impact here. And if these charts make you think about memory usage, then keep in mind that a lot of games like GTA 4, Oblivion, and Fallout 3 have an almost open environment that loads on the fly. A multi-core CPU can permanently and constantly speed up this process, leading to a smoother experience. In RTS games like StarCraft II and Supreme Commander 2, frame rate drops due to the loading of new effects are less severe.

One should always investigate minimum frame rates. What good is new and expensive hardware and high average frame rates if you experience dips down to unplayable levels? The results below only take into account the smallest value in a long sequence of numbers, and the sequences are not completely identical.

Note that minimum values only represent a small snapshot of the benchmark. They don't indicate the length of the frame rate drop, the frequency, or the cause. That makes the method of measurement one of limited use, and it should only be used as a supplement for some of the benchmarks in this article.

The latest 3D engines aren't doing Intel or AMD any favors. Somehow, something seems a bit off here. When most games used DirectX 9, you needed a really fast CPU, and these weren't terribly common before the 65 nm fab node due to thermal issues. Transitions to 65, and then 45 nm allowed processor manufacturers and overclockers to accelerate clock frequencies, making 3.5 GHz a breeze with most CPUs.

But the newer DirectX 10 and 11 games don't really need this speed. This change snuck up on us, making CPU overclocking for 3D games rather meaningless at the mainstream level. Of the 20 games tested in this article, only 10 respond at all to CPU overclocking when we used a GeForce GTX 460 graphics card. Seven of them show only small reactions. Just three show a small frame rate increase. Now, this situation might be completely different at lower resolutions, but the Radeon HD 5800-class monitors, the GeForce GTX 460, and HD resolution LCD monitors are slowly becoming standard gaming equipment.

The following table provides an overview of all the test results so that you can check whether your PC is ready for the tested games or if you need to upgrade something. If you are looking for an upgrade, we show what you should replace first. All of the games deliver higher frame rates with better graphics cards. The average optimal number of CPU cores suggested by the test results is 2.75, showing a clear trend towards at least three CPU cores. The age of dual-core CPUs is certainly not over yet, since DirectX 11 yet again puts the focus on graphics card performance. But with the help of better supply and pricing, the market is trending toward quad-core CPUs nonetheless (or at least dual-core chips with Hyper-Threading). Overclocking the CPU is quite pointless when playing games with high quality graphics settings and HD resolutions--unless you already own a pair of flagship graphics cards rendering cooperatively able to shift the bottleneck back to the host processor.

Games and Impact Factors
GPU Performance Impact
Recommended Number of CPU Cores
Game Response to CPU Overclocking
Alien vs. Predatoryes1no
Alpha Protocolyes2no
Anno 1404yes2no
Battlefield: Bad Company 2yes4somewhat
Bioshock 2yes2somewhat
Call of Duty: Modern Warfare 2yes2somewhat
Divitiny 2: Ego Draconisyes2no
Dragon Age: Originsyes4somewhat
GTA 4 EFLCyes4yes
Just Cause 2yes2no
Kane & Lynch 2yes2no
Left 4 Dead 2yes4yes
Mass Effect 2yes4somewhat
Metro 2033yes2no
Prince of Persia: Forgotten Sandsyes2somewhat
Stalker Call of Pripyatyes2somewhat
StarCraft IIyes4no
Supreme Commander 2yes4yes

Giving a straight answer regarding the amount of graphics memory you need is difficult. The games adapt. If less memory is available, less is often used. The feeling we get is that the 768 MB of our Nvidia GeForce GTX 460 graphics card is already starting to feel somewhat insufficient, although it's still fine for most games, so long as you aren't cranking up the anti-aliasing. At a 1920x1200 resolution, 8xAA, and 16xAF in a DirectX 11 game, you should aim for at least 1 GB of graphics memory.

The question of whether the CPU or GPU is most important is easily answered. If you don't have a multi-core CPU, then upgrade it. If you have a dual-core CPU at around 3 GHz, then invest your money into a graphics card, as most games are GPU-limited. This is not something that will change with new DirectX 11 games.

No comments:

Post a Comment