Basemark's Cross-Platform GPU Test: iPhone 14 Pro Beats Steam Deck

Basemark introduced its new GPUScore Sacred Path cross-platform benchmark that lets users compare completely different GPU architectures running on different operating systems. Our GPU benchmarks hierarchy is limited to Windows 11 testing, but what if you want to know how the best graphics cards stack up against your smartphone? Now you can.

To give a few examples, Sacred Path shows how Apple's A16 GPU stacks up against Steam Deck's GPU and Samsung's Galaxy S22+ GPU as well as how performance of Nvidia's GeForce RTX 3090 relates to that of Intel's Arc A770 and Apple's M1 Max. Of course, with any benchmark that has to work across a wide spread of hardware capabilities, there are caveats. But let's see what the numbers say.

Cross-Platform, Cross-Device Benchmark

There are quite a few major hardware graphics platforms for PCs and mobile devices. These include AMD's integrated and discrete GPUs, Apple's PowerVR-derived integrated graphics processors, Arm's Mali graphics solutions, Imagination's PowerVR-based graphics solutions, Intel's standalone Arc Alchemist and integrated Xe-LP GPUs, Nvidia's discrete graphics processors, Qualcomm's Adreno graphics for mobile applications, Samsung RDNA 2-based built-in GPU cores, and a number of other platforms.

Formally, all of these GPUs are designed to enable high-performance gaming, though high-end mobile and high-end PC are obviously quite different targets. All the architectures have their own pros and cons, and in many cases they run under different operating systems using different APIs (application programming interfaces). That makes it very difficult to determine how one GPU stacks up against another, since it is almost impossible to do an apples-to-apples — or Apples-to-desktops — comparison.

Basemark's cross-platform Sacred Path is designed to clear away some of the confusion, as it performs exactly the same workload on all the various GPUs (assuming the test is run using the official native resolution of 2560x1440). To make the benchmark more relevant for different devices, it supports various performance optimization techniques, including variable rate shading, advanced custom shaders, Ground Truth Ambient Occlusion (GTAO), and temporal anti-aliasing.

Repeatable, but With Fluctuations

After playing with the Sacred Path benchmark for a couple of hours on several Apple devices, we found that while in general the program produces consistent and repeatable results that correlate with the GPU capabilities of the device, at times its results fluctuate significantly — we're talking ~50% or more here. Perhaps Apple's mobile platforms adapt to workloads to provide better performance, but since smartphones and tablets are not in our wheelhouse, we'll leave appropriate investigation to someone else.

Basemark's PowerBoard 4.0 database already has enough submissions at the official render resolution of 2560x2440 to allow comparisons among many different platforms. To provide a better perspective, we compared the highest results each graphics processor achieved and did not take the CPU or API into account.

Unfortunately, since these results can fluctuate quite a bit, we'll have to apply appropriate levels of salt to the PowerBoard 4.0 submissions. Also keep in mind that the results from smartphones, notebooks, and tablets may vary greatly due to different cooling and ambient temperatures — these are not all run under test lab conditions by any stretch of the imagination.

High-End Gaming GPUs

Basemark's Sacred Path benchmark doesn't look particularly amazing compared to modern games, but it still needs a lot of resources to run at decent speeds. Sacred Path's score represents the average FPS multiplied by 100 (e.g, 1840 points means 18.4 FPS). Fortunately, even existing GPUs from AMD and Nvidia can crack this benchmark pretty easily, delivering 200–280 FPS. That also means they're potentially not getting anywhere close to 100% GPU utilization, which again is the problem with trying to create a "universal" GPU benchmark.

Judging by results demonstrated by high-end gaming graphics cards, we can conclude that Basemark's Sacred Path loves raw compute performance. The more FP32 teraflops you have, the better your results will be. AMD's Radeon RX 6900 XT is very close to the GeForce RTX 3080 Ti, but the GeForce RTX 3090 is unrivalled.

Sadly, there are no RTX 4090 submissions just yet. Jarred ran it but it failed to submit (perhaps the system currently has a whitelist of known GPUs and the 4090 isn't on it). Manual (custom) testing using the official settings gave a result of 48,709 over the ~4 second run, which at least seems to be in the right ballpark. At the same time, a test that completes in seconds means that the GPU isn't even remotely warmed up before the test is over. (Where's my salt shaker?)

Mid-Range and High-End Mobile Gaming GPUs

Midrange and high-end mobile gaming graphics processors don't have as much oomph as high-end desktop cards. Nonetheless, they all — AMD's Navi 22, Intel's ACM-G10, and Nvidia's GA104 — do quite well. Intel's Arc A770 is 11% behind Nvidia's GeForce RTX 3060 Ti, as another point of reference.

Meanwhile, assuming that Apple's M1 Ultra scales its graphics performance really well (the M1 Ultra consists of two M1 Max SoCs connected together), we would expect its performance to be on par with that of Nvidia's GeForce RTX 3070, which is not bad for an integrated graphics processor.

Entry-Level GPUs

Owners of entry-level GPUs have not submitted many results so far, or at least they're not showing up. Fundamentally, the more teraflops a GPU has, the better it performs in Sacred Path.

Apple's M1 Max (an iGPU) and Nvidia's GA106 in the form of desktop GeForce RTX 3050 lead the pack here. Meanwhile, Intel's standalone Arc A380 can barely keep up with Apple's M1 Pro integrated GPU. Even Nvidia's GeForce RTX 3050 for laptops is significantly faster than the Intel Arc A380.

Handhelds, Ultra-Portables, Integrated

When it comes to cross-platform benchmarking, handhelds, ultra-portables, and iGPUs are arguably the most interesting categories as they includes multiple hardware and software architectures as well as different form-factors.

Since Apple uses its M1 and M2 system-on-chips for its tablets, it's reasonable to include them into this comparison even though they can power full-fledged PCs like MacBook Air/MacBook Pro. After all, AMD's Ryzen SoCs with higher-end GPUs are also designed for this very purpose. Meanwhile, Apple's M1 and M2 iGPUs are faster than AMD's Cezanne iGPU by 40%–80%. 

As for SoCs for handheld devices, Apple's A16 can beat even the RDNA 2-based AMD Van Gogh iGPU used in the Steam Deck game console, but the iPhone 14 Pro needs to run the test for a couple of times to hit something like 2200 points. Qualcomm's Adreno 730 used in the Snapdragon 8 Gen 1 (SM8450) is a good performer, but it's still behind Apple's previous-generation A15 SoC. Meanwhile, RDNA 2-based Samsung Xclipse 920 looks somewhat disappointing when compared to rivals.

Summary

Basemark's Sacred Path benchmark can indeed be used to test all kinds of graphics processors, from humble GPUs aimed at smartphones and all the way to monstrous graphics cards that can be used not just for games but also as space heaters or AI supercomputers.

The software produces consistent and repeatable results most of the time, and it heavily depends on the compute performance of modern GPUs. While that might be useful on some level as a performance metric, do keep in mind that a 1.15GB benchmark that can complete in seconds on the fastest GPUs will never adequately show the differences between GPUs that might have anywhere from 1GB to 24GB (or more) VRAM. In this test, raw teraflops wins the game, even though actual game titles also depend on graphics memory capacity, memory bandwidth, and other factors.

The benchmark also doesn't use ray tracing in a bid to work on all graphics processors from the past decade or more, so it cannot demonstrate all the benefits that leading-edge GPUs for PCs support. There are also some fluctuation-related hiccups with some platforms, but that's enough for this look at the vagaries of cross-platform "universal" GPU testing.