GPU And Gaming Performance
In the previous section, we saw Snapdragon 820 excel in the sequential memory bandwidth tests, which should help boost overall GPU performance. With memory bandwidth a less likely bottleneck, will a weak point emerge in the Adreno 530’s architecture? Will it show an increase in ALU performance like we expect?
In 3DMark Ice Storm Unlimited, the Adreno 530 in Snapdragon 820 falls behind the PowerVR GT7600 GPU in the A9 and even the older Adreno 430 in Snapdragon 810. Curiously, the Adreno 530 trails the 430 in both pixel- and vertex-heavy workloads. As we will see, however, these results are an outlier. Because of the Snapdragon 820 MDP’s pre-production status, this could simply be a quirk of an early revision graphics driver.
The 3DMark Physics test focuses on CPU performance and, like the AndEBench Memory Latency test, uses a data structure with randomized elements. This is why we see all of Qualcomm’s most recent SoCs, including the 808, 810, and 820, along with Apple’s A9, perform poorly. Their memory controllers are optimized for orderly, sequential memory access.
Snapdragon 820 fares better in Basemark X, with a 29% advantage over the Snapdragon 810 and a slim 7% advantage over the Exynos 7420. Most of the Adreno 530’s advantage comes from the Dunes test, which is heavily skewed towards vertex processing. Historically, Qualcomm’s Adreno GPUs have always been stronger with pixel operations, with vertex processing a relative weak point compared to other GPUs. It looks like the Adreno 530 is a more balanced design.
If you’re wondering why Apple’s A9 is not included here, it’s because Basemark X will not run on iOS 9 due to some API changes.
At the high quality setting, Snapdragon 820 flexes its memory bandwidth muscles and extends its performance lead to 57% over Snapdragon 810 and 30% over the Exynos 7420. Even in the onscreen tests, where the Snapdragon 820 is processing more pixels than any other device in these charts, the Adreno 530 performs well, which makes it suitable for the high-resolution QHD screens common in smartphones today.
We ran two different versions of the GFXBench Manhattan test. The older Manhattan 3.0 version runs an OpenGL ES 3.0 based game engine, with an abundance of lighting and pixel effects. Looking at the offscreen results we see that the Adreno 530 doubles the performance of both the Adreno 430 in Snapdragon 810 and the ARM Mali-T760MP8 in Exynos 7420. It’s even 20% faster than Apple’s A9.
Manhattan 3.0 uses deferred rendering for its lighting effects, which is highly dependent on screen resolution. This is why the iPhone 6s Plus (A9) and OnePlus 2 (Snapdragon 810) move up the chart when rendering onscreen: both have native 1920x1080 displays compared to the 2560x1600 display on the Snapdragon 820 MDP.
Manhattan 3.1 is an enhanced version of the Manhattan test utilizing OpenGL ES 3.1 features. Snapdragon 820’s performance advantage in this test is quite impressive; it’s 79% faster than the Snapdragon 810 and 2.4x faster than the Exynos 7420.
While the Manhattan tests emphasize pixel shading, the OpenGL ES 2.0 based T-Rex game simulation uses a more balanced rendering pipeline. Once again the 820 offers an impressive boost over the 810, this time around 77%. The 820 even beats Apple’s best offering by 14%, which held a stranglehold on T-Rex for some time.
The GFXBench Alpha Blending test stresses rasterization and memory bandwidth. Given the 820’s excellent memory performance, it’s no surprise to see it rise to the occasion, beating Apple’s A9 by 25%. The Fill test also stresses memory bandwidth, in addition to pixel processing, both of which play to the 820’s strengths.
Qualcomm has been improving ALU performance in its Adreno GPUs for several generations now, and the Adreno 530 is no exception. Unfortunately, we cannot say for sure if the 530 includes additional ALUs or even estimate total GFLOPS, because we cannot separate the individual contributions of the ALUs and memory bandwidth from their aggregate performance. All we can say is that Snapdragon 820 delivers 43% better ALU performance than the 810 and that a good chunk of this likely comes from the increase in memory bandwidth. ALU performance is critical to Qualcomm’s heterogeneous computing vision, and it looks like the Snapdragon 820 is well equipped to execute it.
