Snapdragon X Elite lags behind iPhone 15 Pro in GravityMark GPU test in a shocking anomaly

The newly launched Snapdragon X Elite chip has finally started shipping in laptops. As per our in-depth review of the VivoBook S 15 OLED featuring the X Elite (78-100) with 12 cores and slightly less powerful graphics than the highest-end SKU (84-100), the chip boasts decent performance and efficiency, although not quite good enough to meet the expectations set by the hype.

In a recent post on Twitter, however, a user reported a GravityMark benchmark score for the X Elite's (78-100) Adreno GPU, where the chip managed to score a measly 20.5 FPS - even lower than that of the A17 Pro chip found in the iPhone 15 Pro which scored 20.9 FPS. In the same test, the Apple M3, which is in the same price bracket as the X Elite, scored around 45.1 FPS.

This result is almost certainly an outlier, as more mainstream graphics benchmarks, such as GFXBench, show the 3.8 TFLOPS Adreno GPU in the X Elite (78-100) handily outperforming the A17 Pro by around 100% and reaching surprisingly close to the 10-core Apple M3 GPU.

It is not clear why the X Elite performed so poorly in this specific test, especially considering the fact that GravityMark has cross-platform support, with a native ARM version for Windows. However, a performance hit might be caused by the DirectX graphics API used for the test. The iPhone 15 Pro used Apple's in-house Metal API which has a lower overhead and more direct access to the GPU, thanks to Apple's tight hardware-software integration. DirectX on the other hand, was designed for a wide range of different hardware configurations, prioritizing versatility above all. The Vulkan API has a lower overhead than DirectX, but was not used in this specific test and might've resulted in a better score. 

However, even an API bottleneck would not be able to adequately explain the weighty performance drop that the Adreno GPU witnessed in this particular benchmark. Perhaps most likely, this is an optimization issue on GravityMark's end, which will assumably be resolved in future versions of the software. So until then, it is best not to judge the new kid on the block based on a single benchmark run.