Crysis 3 Performance, Benchmarked On 16 Graphics Cards

Test System And Graphics Hardware

As always, we strive to represent game performance across a wide range of graphics hardware. We're including cards from the low-end Radeon HD 6450 and GeForce GT 630 GDDR5 to the GeForce GTX 690, including multi-card Radeon HD 7970 CrossFire setups (note that we also wanted to include the low-end GeForce 210, but it's not DirectX 11-compatible and wouldn't run the game).

You'll also find benchmark results at 5760x1080, in case you're interested in Surround or Eyefinity performance.

We want to mention that PowerColor supplied us with a test sample of the new Radeon HD 7870 LE for this performance analysis. Unlike the Pitcairn-based Radeon HD 7870, the Radeon HD 7870 LE is built with a cut-down Tahiti GPU from the Radeon HD 7900 series.  The card offers 1,536 shaders and 1,500 MHz memory, and while it can be purchased for the same $240 as a garden-variety Radeon HD 7870, it performs closer to the $300 Radeon HD 7950.

Testing Notes

Our goal is to represent worst-case real-world performance. There is no point in benchmarking an area of the game that subjects your system to a relatively light load. If we based our hardware recommendations on that, you'd make your buying decision based on our tests and then run into problematic frame rates in more demanding sections of the game. So, we sought out the most challenging sequence possible, loaded with enemies that'd exact maximum load from the CPU and GPU.

We ended up choosing an area in the vast "Welcome to the Jungle" level. It includes a lot of lush foliage, soldiers, and things going on to challenge our test system. We often see comments in the forums that people with a similar hardware setup experience higher frame rates than our benchmarks: that's because they are testing a lighter load than we are.

We're including our new consecutive frame time metric, but we've modified the formula a little. Instead of simply reporting the variance between consecutive frames, we're comparing the variance between each frame and an ideal frame time calculated based on a range of previous and following frames. We'll go into this more in an upcoming story dedicated to our efforts. But this evolving approach removes some of the confusion with our results and natural variances in the frame rate. Note that we've also changed the nomenclature from 'latency' to 'variance' in order to more accurately describe our focus.

Keep in mind that we set all overclocked cards to reference specifications to represent the majority of products on the market.