Going Beyond Performance Testing
Graphics standards continue to evolve, and benchmarks are frequently updated to keep pace. In many ways, GFXBench 3.0 is a new beast, featuring the suite's first OpenGL ES 3.0-based performance test, measures of image quality, system impact, and battery performance.
Previous GFXBench versions were comprised of traditional “high-level” game-like scenarios along with more “low-level” tests designed to measure specific subsystems. Version 3.0 expands the software's scope, retaining only one sequence from previous builds: the OpenGL ES 2.0-level T-Rex HD from GFXBench v2.7. The old faithful Egypt HD test from v2.5 retires, which makes sense, since even modern entry-level graphics engines power through it with ease. Naturally, it's no longer challenging for mid-range and high-end SoCs. In its place is the far more demanding Manhattan test, utilizing OpenGL ES 3.0-specific complex lighting, particles, and, most important, deferred shading.
The Low Level performance benchmarks have also been improved over v2.7. An ALU test was added to calculate raw shader performance, while the new Alpha Blending test does the same for rendering multiple transparent objects on top of each other. In addition, there’s a new set of render quality tests that evaluate a device's fidelity by comparing a single rendered frame to a reference, scoring the outcome in peak signal-to-noise ratio. One version forces the shaders to run with high precision, and the other doesn't. A Driver Overhead test shows how heavily a CPU complex is affected by draw calls and state changes.
Finally, a brand new battery test was added, rendering the T-Rex HD in a loop at 50% screen brightness and logging frame rate as the test iterates at least 30 times. While Manhattan may get all the spotlight for emphasizing current-gen graphics capabilities, this is also a very important addition. It, along with the quality metric, gives us a way to compare performance to output and longevity. Device makers and SoC vendors who optimize for one vector are going to negatively affect the others.
Today we’ll give you a test-by-test analysis of GFXBench 3.0 using a selection of devices that run the gamut of modern SoCs.
|Device||SoC||CPU Core||GPU Core||Memory||Display||Battery|
|Apple iPhone 5s||Apple A7||ARM v8 (dual-core) @ 1.3 GHz||Imagination Technologies PowerVR G6430 (four-cluster) @ 300 MHz||1 GB DDR3||4" IPS @ 1136x640 (326 PPI)||1560 mAh|
|EVGA Tegra Note 7||Nvidia Tegra 4 (T114)||ARM Cortex-A15 (quad-core) @ 1.8 GHzARM Cortex -A15 (single companion-core) @ 500 MHz||GeForce ULP (72-core) @ 672 MHz||1 GB DDR3||7" IPS @ 1280x800 (216 PPI)||4100 mAh|
|Google Nexus 5||Qualcomm Snapdragon 800 (8974-AA)||Qualcomm Krait 400 (quad-core) @ 2.3 GHz||Qualcomm Adreno 330 (quad-core) @ 450 MHz||2 GB DDR3||4.94” IPS+ @ 1920x1080 (445 PPI)||2300 mAh|
|Google Nexus 7||Qualcomm Snapdragon S4 Pro (APQ8064-1AA)||Qualcomm Krait 300 (quad-core) @ 1.5 GHz||Qualcomm Adreno 320 (quad-core) @ 400 MHz||2 GB DDR3||7.1” IPS @ 1920x1200 (323 PPI)||3950 mAh|
|Meizu MX3||Samsung Exynos 5 Octa (5410)||ARM Cortex-A15 (quad-core) @ 1.6 GHzARM Cortex-A7 (quad-core) @ 1.2 GHz||Imagination Technologies PowerVR SGX544MP3 (tri-core) @ 480 MHz||2 GB DDR3||5.1” IPS @ 1800x1080 (412 PPI)||2400 mAh (Li-Pro)|
|Oppo N1||Qualcomm Snapdragon 600 (APQ8064T)||Qualcomm Krait 300 (quad-core) @ 1.7 GHz||Qualcomm Adreno 320 (quad-core) @ 400 MHz||2 GB DDR3||5.9” IPS @ 1920x1080 (373 PPI)||3610 mAh|
|Samsung Galaxy Note 10.1" 2014 Edition||Samsung Exynos 5 Octa (5420)||ARM Cortex-A15 (quad-core) @ 1.9 GHzARM Cortex-A7 (quad-core) @ 1.2 GHz||ARM Mali-T628MP6 (hexa-core) @ 480-600 MHz||3 GB DDR3||10.1” WQXGA TFT @ 2560x1600 (229 PPI)||8220 mAh|
Let’s start out with Manhattan, GFXBench’s newest High-Level test, and the first serious OpenGL ES 3.0 benchmark to hit the scene.