PhysX Performance Update: GPU vs. PPU vs. CPU August 14, 2008 Brandon Sandman Bell

Introduction

Page:: ( 1 / 2 )

In those tests we generally found that PhysX apps ran around three times faster on the GPU in comparison to the CPU, although the exact amount varied depending on the game being tested – some cases were up to four times faster.

After seeing the results, we received numerous requests from readers asking us to include scores with AGEIA’s PhysX PPU. Many of you were curious to see how the PPU stacked up to today’s GPUs when it came to PhysX performance.

To accomplish this we dusted off our old PhysX card manufactured by BFG, and plopped it into an open Core 2 QX9650 testbed we’d already had setup. Surprisingly enough, the board still ran like a charm! We then proceeded to install the PhysX apps and came up with another custom demo for Unreal Tournament 3 testing taken on the heat ray map under similar conditions (our original demo used for testing was deleted when we wiped the testbed’s hard drive). Once again we used the gravity gun’s secondary fire mode exclusively to send objects flying throughout the map. Our tests with Warmonger and Nurien use the standard built-in benchmarks that ship with each app (we actually use FRAPS to record Nurien scores, as the built-in fps results are inaccurate). The MKZ alpha doesn’t seem to work with the PPU at the moment. Every time we tried to run the benchmark, the game crashed to the Windows desktop.

So how did the PPU fare against the GPU? First let’s take a look at UT3:



As you can see, going from the CPU to the PPU netted us a gain of roughly 2.5X. All of the cards generally top out around 32-34 frames per second with the PPU handling physics, versus 13 fps with CPU-based PhysX. The GeForce GPUs then run 16% faster than PPU-based PhysX in UT3.

Nurien

If you recall our original article, the cloth simulation in Nurien wasn’t as demanding as the physics calculations in the other games we tested, and thus we saw slimmer gains in favor of GPU-based PhysX. Based on this, we expected PPU performance to be at least on par with GPU PhysX performance, if not better since we’re running dedicated PhysX hardware. Surprisingly enough however this wasn’t the case:



We re-ran the tests multiple times to ensure that our PPU results were accurate. Each time though the scores remained pretty consistent. The PPU just wasn’t scaling beyond 36 fps, an improvement of 1.5 times for the GeForce boards over the CPU, while the Radeon cards ran about 14% faster with the aid of the PPU.

The GeForce PhysX setup still ran over 20% faster than the PPU setups though.

Of course, Nurien is pre-release alpha software. It was probably never designed with AGEIA’s PPU in mind. This isn’t the case for our next app, Warmonger…

Warmonger/Conclusion

Page:: ( 2 / 2 )

The last PhysX app we’re testing is Netdevil’s FPS Warmonger. Warmonger uses PhysX extensively, besides Cell Factor: Revolution, it was the first full game that was truly designed from the ground up for the PhysX physics processing unit. Let’s see how this game runs with GPU vs. PPU vs. CPU-based PhysX:



Like UT3, running PhysX on the CPU severely hampers Warmonger performance. The GeForce cards are capped at around 15-16 fps, while the Radeon boards managed to hit frame rates in the 19 fps range. As expected, the PPU and GPU perform significantly better. PPU-based PhysX actually runs very close to the GPU-based PhysX cards. In the case of the GeForce 8800 GT, the PPU only trails the GPU by 12%. With its extra stream processors and higher clocks the 9800 GTX is able to pull away slightly further, 16%.

Conclusion

For hardware that’s over two years old, AGEIA’s PhysX card was able to put up a much stronger showing than we expected, but it still wasn’t able to match the performance of today’s GeForce 8800 GT and 9800 GTX cards. In UT3 for instance the 8800 GT was 16% faster than the PPU (the closest example was Warmonger at 12%). And keep in mind that the GeForce cards we tested were running a mixture of 3D and PhysX, while the PPU is dedicated solely to PhysX calculations. In the original PhysX performance article we saw some pretty nice gains in Warmonger and UT3 when one GeForce 9600 GT card was devoted entirely to running PhysX.

Based on all this, we stand by the conclusions we drew in our original GeForce PhysX performance article. With PhysX cards selling for $120-$125 on TigerDirect and Newegg it just doesn’t make sense to purchase a dedicated PhysX card when you can buy a GeForce 8800 GT for the same amount of money. You’ll get better performance and a card that can be used for more than just PhysX.

Ultimately though we’re still waiting for more engaging PhysX content to be released. PhysX is a nice little bonus worth checking out if you already own a GeForce 8 or better GPU with 256MB of graphics memory, but as it stands now it isn’t compelling enough to make Radeon 4800 card owners jealous in our opinion.

Considering NVIDIA’s strong developer relations, it’s possible that in time this could change. Now that PhysX is supported by millions of GeForce boards, developers may consider taking a second look at the technology. It remains to be seen when or even if that will happen though.