NVIDIA APEX PhysX Efficiency: CPU vs GPU

Benchmark Reviews tests NVIDIA APEX PhysX efficiency using Mafia II - compares CPU vs GPU performance.

According to the August 2010 Steam hardware survey, PC gamers are using NVIDIA GeForce desktop video cards nearly 80% more than AMD/ATI counterparts. Great products have come from both GeForce and Radeon brands, yet based on this survey NVIDIA owns almost 60% of the entire graphics market compared to AMD's 33%. Gamers might rely on NVIDIA's hardware for its superior graphical processing power and affordable price point, but it's their gaming technologies that have helped deliver complete market dominance (among Steam users). NVIDIA's "The Way It's Meant to be Played" is a trademarked slogan denoting a direct involvement in software development as much as they focus on hardware. When the Ageia PhysX software physics technology was purchased back in early 2008, that commitment sharpened NVIDIA's growing double-edge sword. Adding 3D Vision only helped consummate their efforts.

In this article, Benchmark Reviews will demonstrate how far PhysX technology has come using the recently-released Mafia-II video game by 2K Games. In this single-player third-person action shooter developed by 2K Czech for 2K Games, players assume the life of World War II veteran Vito Scaletta, the son of small Sicilian family who immigrates to Empire Bay. Mafia II makes use of DirectX-11 extensions on 2K Czech's proprietary Illusion game engine, which introduces NVIDIA APEX PhysX and GeForce 3D-Vision technology enhancements. NVIDIA's APEX PhysX modeling engine adds new Destruction, Clothing, Vegetation, and Turbulence physics into games such as Mafia II. While adding PhysX support to a video game is nothing new for NVIDIA, allowing APEX PhysX features to be computed by the computer's central processor is new territory. For this NVIDIA APEX PhysX: CPU vs GPU Efficiency demonstration, our tests compare GeForce and Radeon GPU's against the Intel Core i7 CPU.

This article isn't intended to become a NVIDIA vs AMD topic, but it becomes impossible to avoid since ATI does not license PhysX. NVIDIA offers a free software development kit so CUDA drivers can be built for AMD products, yet all ATI Radeon graphics cards (up to the HD 5000 series) still do not compute PhysX commands without using modified drivers. As a result, PhysX hardware acceleration is presently available only on GeForce GPUs unless gamers research unsupported options for their Radeon products. NVIDIA has opened their PhysX platform to AMD and Intel processors in Mafia II, allowing hardware acceleration to be calculated my the system's central processor. The narrative of this article is how well PhysX is processed by the CPU and GPU, and where the different GeForce Fermi graphics processors (GF100, GF104, GF106) stack up in regards to PhysX efficiency.

NVIDIA APEX PhysX Destruction, Clothing, and Particles in Mafia II

Mafia II is the sequel to Mafia: The City of Lost Heaven released in 2002. Growing up in the slums of Empire Bay teaches Vito about crime, and he's forced to join the Army in lieu of jail time. After sustaining wounds in the war, Vito returns home and quickly finds trouble as he again partners with his childhood friend and accomplice Joe Barbaro. Vito and Joe combine their passion for fame and riches to take on the city, and work their way to the top in Mafia II. While this premise makes for an interesting storyline, it's the graphical effects that keep players immersed in very realistic settings. NVIDIA's APEX PhysX is the glue that binds in Mafia II, giving the game life-like physics to create a virtual reality.

Mafia II was developed with NVIDIA's PhysX version 2.8.3, the available software development kit at the time. This version supports only single-core single-threaded PhysX CPU processing, which is minimal in comparison to the available hardware of most PCs. The current PhysX SDK (version 2.8.4) supports SSE2 instructions, but this feature must be enabled by the developer. According to NVIDIA, the forthcoming PhysX SDK 3.0 is said to introduce multi-threaded CPU support for PhysX extensions, and SSE is enabled by default.

NVIDIA APEX PhysX Enhancements

PhysX helps make object movement more fluid and lifelike, such as cloth and debris. Mafia II is the first PC video game title to include the new NVIDIA APEX PhysX framework, a powerful feature set of graphical special effects that only GeForce video cards can currently deliver. Only the PC version of Mafia II supports NVIDIA's APEX PhysX physics modeling engine, which adds the following features: APEX Destruction, APEX Clothing, APEX Vegetation, and APEX Turbulence. This section demonstrates the differences between playing Mafia II with and without APEX PhysX enabled.

Many of the latest video games are being developed with new graphical enhancement technologies, but Mafia II is the first to offer both NVIDIA APEX PhysX and 3D-Vision Surround features together. Each of these NVIDIA technologies are designed to work their best on GeForce desktop graphics solutions, but only the most powerful GPUs can make these high-quality special effects stand out in full glory. We begin with a scene from the Mafia II benchmark test, which has the player pinned down behind a brick column as the enemy shoots at him. Examine the image below, which was taken with a Radeon HD 5870 configured with all settings turned to their highest and APEX PhysX support disabled:

No PhysX = Cloth Blending and Missing Debris

Notice from the image above that when PhysX is disabled there is no broken stone debris on the ground. Cloth mesh from the foreground character's trench coat blends into his leg and remains in a static position relative to his body, as does the clothing on other (AI) characters. Now inspect the image below, which uses the GeForce GTX 480 with APEX PhysX enabled on high:

PhysX Enabled = Realistic Cloth Mesh and Falling Debris

With APEX PhysX enabled, the cloth vertices neatly sway with the contour of a characters body, and doesn't bleed into solid objects such as body parts. Additionally, APEX Clothing features improve realism by adding gravity and wind effects onto clothing, allowing for characters to look like they would in similar real-world environments. This added realism makes a noticeable difference during game play.

Burning Destruction Smoke and Vapor Realism

Flames aren't exactly new to video games, but smoke plumes and heat vapor that mimic realistic movement have never looked as lifelike as they do with APEX Turbulence. Fire and explosions added into a destructible environment are a potent combination for virtual-world mayhem, showcasing the new PhysX APEX Destruction feature.

Exploding Glass Shards and Bursting Flames

NVIDIA PhysX has changed video game explosions into something worthy of cinema-level special effects. Bursting windows explode into several unique shards of glass, and destroyed crates bust into splintered kindling. Smoke swirls and moves as if there's an actual air current blowing, and flames move out towards open space all on their own. Surprisingly, there is very little impact on FPS performance with APEX PhysX enabled on GeForce video cards, and very little penalty for changing from medium (normal) to high settings. We prove this in the next section...

Testing and Initial Results

VGA Testing Methodology

The Microsoft DirectX-11 graphics API is native to the Microsoft Windows 7 Operating System, and will be the primary O/S for our test platform. DX11 is also available as a Microsoft Update for the Windows Vista O/S, so our test results apply to both versions of the Operating System. The majority of benchmark tests used in this article are comparative to DX11 performance, however some high-demand DX10 tests have also been included.

According to the Steam Hardware Survey published for the month ending August 2010, the most popular gaming resolution is 1280x1024 (17-19" standard LCD monitors). However, because this 1.31MP resolution is considered 'low' by most standards, our benchmark performance tests concentrate on higher-demand resolutions: 2.30MP 1920x1200 (24-28" widescreen LCD monitors). These resolutions are more likely to be used by high-end graphics solutions, such as those tested in this article.

In each benchmark test there is one 'cache run' that is conducted, followed by five recorded test runs. Results are collected at each setting with the highest and lowest results discarded. The remaining three results are averaged, and displayed in the performance charts.

Intel X58-Express Test System

• Motherboard: ASUS P6X58D-E (Intel X58/ICH10R/Marvell SE9128)
• Processor: Intel Core i7-920 Nehalem 2.66 GHz (BX80601920)
• CPU Cooler: Prolimatech Megahalems
• System Memory: Kingston HyperX DDR3 KHX2000C8D3T1K3 (3x2GB @ 1333 CL6-6-6-18)
• Primary Drive: OCZ Agility-EX SLC SSD OCZSSD2-1AGTEX60G firmware 1.60
• Power Supply Unit: OCZ Z-Series Gold 850W OCZZ850
• Monitor: 26-Inch Widescreen LCD (up to 1920x1200@60Hz)

DirectX-11 Benchmark Applications

• Mafia II
  • Extreme Settings: (Antialiasing, 16x AF, High Shadow Quality, High Detail, Ambient Occlusion, High APEX PhysX)

Video Card Test Products

• NVIDIA GeForce GTS 450: 783 MHz GPU/1566 MHz Shader/902 MHz vRAM - Forceware 260.63
• ATI Radeon HD 5770 (CrossFireX): 850 MHz GPU/1200 MHz vRAM - ATI Catalyst Driver 10.8
• NVIDIA GeForce GTX 460 1GB: 675 MHz GPU/1350 MHz Shader/900 MHz vRAM - Forceware 260.63
• ATI Radeon HD 5850: 725 MHz GPU/1000MHz vRAM - ATI Catalyst Driver 10.8
• NVIDIA GeForce GTS 450 1GB (SLI): 783 MHz GPU/1566 MHz Shader/902 MHz vRAM - Forceware 260.63
• ATI Radeon HD 5870: 850 MHz GPU/1200MHz vRAM - ATI Catalyst Driver 10.8
• NVIDIA GeForce GTX 480: 701 MHz GPU/1401 MHz Shader/924 MHz vRAM - Forceware 260.63

PhysX Level Impact

Mafia II offers comprehensive video settings, with everything from antialiasing (on or off) and anisotropic filtering (1/2/4/8/16x) to ambient occlusion (on or off) and APEX PhysX (off/medium/high). Our benchmark tests use all settings configured to their highest quality, and then added medium or high APEX PhysX. Measured at 1920x1200 resolution, Mafia II placed ample demand on all four graphics cards tested. The purpose of this test is to illustrate how much impact PhysX has at three levels: off, medium, and high.

Rather than grading one product against another as we usually do in our video card reviews, these results help us focus on individual impact. Here's an initial reaction to how each video card is affected by PhysX performance (more tests in the next section):

• Radeon HD 5850
  45.8 FPS baseline frame rate performance with APEX PhysX disabled.
  20.3 FPS with medium-quality PhysX on the GPU, 55.7% reduction in performance over the baseline.
  15.3 FPS with high-quality PhysX on the GPU, 66.7% decrease over the baseline and 24.6% reduction in performance over medium quality.
• GeForce GTX 460 1GB
  • 33.4 FPS baseline frame rate performance with APEX PhysX disabled.
  • 27.3 FPS with medium-quality PhysX on the GPU, 18.3% reduction in performance over the baseline.
  • 25.1 FPS with high-quality PhysX on the GPU, 24.9% decrease over the baseline and 8% reduction in performance over medium quality.
• Radeon HD 5870
  • 53.3 FPS baseline frame rate performance with APEX PhysX disabled.
  • 21.5 FPS with medium-quality PhysX on the GPU, 59.7% reduction in performance over the baseline.
  • 15.7 FPS with high-quality PhysX on the GPU, 70.5% decrease over the baseline and 27% reduction in performance over medium quality.
• GeForce GTX 480
  57.2 FPS baseline frame rate performance with APEX PhysX disabled.
  35.0 FPS with medium-quality PhysX on the GPU, 38.8% reduction in performance over the baseline.
  32.9 FPS with high-quality PhysX on the GPU, 42.5% decrease over the baseline and 6% reduction in performance over medium quality.

What we can surmise from this initial results is that AMD/ATI Radeon HD 5000-series video cards suffer obscene frame rate reductions when APEX PhysX is computed by the CPU, while GeForce products manage the physics models more efficiently. The NVIDIA Fermi GF104 does a much better job of computing PhysX in the GeForce GTX 460 than the GF100 does in the GTX 480. Now let's look at the finer details...

APEX PhysX: GPU vs CPU Tests

At Benchmark Reviews, we test video cards using only the mutually available common settings for each product tested. If our goal was to compare video frame rate performance, we would ensure that anti-aliasing does not exceed 8x MSAA (maximum supported by Radeon GPUs) and PhysX capabilities were disabled (currently only supported on GeForce GPUs). This section isn't about comparing relative graphical processing power between competing products, but instead it focuses on how relevant GPU processing of physics models has become. Intel and AMD have battled for processor supremacy for years, but when it comes to video games NVIDIA has proven that the GPU trumps all.

In this second series of tests, six different video card combinations are tested in Mafia II using disabled, medium, and high APEX PhysX settings. Pay careful attention to how CPU-bound PhysX performance declines in reverse order of performance:

• GeForce GTS 450
  • 23.2 FPS baseline frame rate performance with APEX PhysX disabled.
  • 19.1 FPS with high-quality PhysX on the GPU, 17.7% reduction in performance over the baseline.
  • 18.8 FPS with high-quality PhysX on the CPU, 19% decrease over the baseline and 1.6% reduction in performance over GPU PhysX.
• GeForce GTX 460
  • 36.5 FPS baseline frame rate performance with APEX PhysX disabled.
  • 26.5 FPS with medium-quality PhysX on the GPU, 27.4% reduction in performance over the baseline.
  • 15.1 FPS with high-quality PhysX on the CPU, 58.6% decrease over the baseline and 43% reduction in performance over GPU PhysX.
• Radeon HD 5770 CrossFireX
  • 52.7 FPS baseline frame rate performance with APEX PhysX disabled.
  • PhysX is not able to be processed on the GPU; all PhysX is processed by CPU.
  • 16.5 FPS with high-quality PhysX on the CPU, 68.7% decrease over the baseline.
• GeForce GTS 450 SLI
  • 52.3 FPS baseline frame rate performance with APEX PhysX disabled.
  • 28.3 FPS with medium-quality PhysX on the GPU, 45.9% reduction in performance over the baseline.
  • 15.8 FPS with high-quality PhysX on the CPU, 69.8% decrease over the baseline and 44.2% reduction in performance over GPU PhysX..
• Radeon HD 5870
  • 53.5 FPS baseline frame rate performance with APEX PhysX disabled.
  • PhysX is not able to be processed on the GPU; all PhysX is processed by CPU.
  • 15.9 FPS with high-quality PhysX on the CPU, 70.3% decrease over the baseline.
• GeForce GTX 480
  • 57.8 FPS baseline frame rate performance with APEX PhysX disabled.
  • 33.3 FPS with medium-quality PhysX on the GPU, 42.4% reduction in performance over the baseline.
  • 14.9 FPS with high-quality PhysX on the CPU, 74.2% decrease over the baseline and 55.3% reduction in performance over GPU PhysX.

Similar to the first set of test results, Radeon video cards suffer poor frame rate speeds when PhysX is enabled. Our Intel Core i7-920 quad-core CPU just doesn't compare to the hundreds of cores available in a graphics processor. Both AMD and NVIDIA products suffer heavily reduced performance when APEX PhysX is processed by the computer's CPU, although there appears to be an unexpected trend: the most powerful GPUs offer the inverse in CPU-processed PhysX performance.

Mafia II is clearly dependant on PhysX technology to deliver the realistic scenery and physics gamers enjoy, and future DX11 video games may follow suit. This could relegate ATI Radeon products to rely on the system's central processor to help compensate, and thereby reduce overall performance when the settings include PhysX. For best results in Mafia II, players are suggested to use NVIDIA GeForce products to properly process PhysX features and produce high-enough frame rate performance to enjoy 3D Vision. You can still play Mafia II and future APEX PhysX-enabled PC video games with your Radeon video card, but as our results show there will be hell to pay.

2K Games designed Mafia II using NVIDIA's PhysX 2.8.3 SDK, which supports only single-threaded PhysX CPU processing. PhysX SDK version 2.8.4 supports SSE2 instructions (which are not enabled by default for backwards compatibility), allowing updated games to compute PhysX more efficiently if developers enable this function. Finally, the forthcoming PhysX SDK 3.0 is said by NVIDIA to introduce multi-threaded CPU support to PhysX with SSE enabled by default, which could really change the game for everyone.

READER FEEDBACK: Do you play Mafia II on a NVIDIA GeForce or AMD Radeon DirectX-11 video card with APEX PhysX enabled? Tell us how it's working for you in the comment area below, and please share your quality settings.