GLBenchmark: iPhone 5 v Galaxy S4

[tinmeigut]

http://www.anandtech.com/show/6914/samsung-galaxy-s-4-review/5

I was just re-reading the GPU performance of the S4 and noticed something I don't understand:

In both the GLBenchmark Fill test and Triangle Texture test, the SGX543MP3 in the iPhone 5 outperformed the Adreno 320 in the S4 whether screen on or off;

But in the Egypt HD test (screen off 1080p), the iPhone 5 performed considerably worse than the S4 (27 fps v 40 fps).

And we know that the Fill and Triangle throughput tests measure theoretical performance and the Egypt test simulates real-life gaming performance. Then the question boils down to this: what made the iPhone 5 superior theoretically but inferior when it comes to real life gaming?

 

[ViRGE]

It's hard to say anything definitive without additional tests, but most likely it's PowerVR's "special sauce". PVR is unique because they're the only architecture that is a tile based deferred renderer. Adreno is tile based, but they don't have the deferred rendering abilities that PowerVR does.

Deferred rendering greatly helps in cutting down on overdraw. So in a fill rate test where overdraw is very likely, their architecture would look very good.

 

[ChronoReverse]

It's really just a matter of synthetics testing a single subsystem whereas when you're rendering a real scene, you need a balance of many factors. So even if you're super fast at a few things, you're entirely bottlenecked by something else.

Adreno is probably just better balanced for rendering a scene.