Nearly two weeks ago, Qualcomm invited tech journalists to Maui for the 2019 Snapdragon Tech Summit. At the event, the company unveiled its latest high-end SoC for mobile devices: the Qualcomm Snapdragon 865 mobile platform. Qualcomm says the new Snapdragon 865 boasts a 25% CPU performance increase and a 20% GPU performance increase over the previous generation Snapdragon 855. Also, the new SoC supports LPDDR5 memory and is manufactured on a newer 7nm process. Qualcomm's latest silicon will make its way to 2020 flagships like the Xiaomi Mi 10, OPPO Find X2, and many other high-end smartphones.

But just how much faster is it than the previous generations? We benchmarked Qualcomm's Snapdragon 865 reference device at the event to find out. We pit the new SoC against the Snapdragon 855+, the Snapdragon 855, the Snapdragon 845, and the Kirin 990 from Huawei's HiSilicon. We would have loved to test the Snapdragon 865 against the MediaTek Dimensity 1000 or Samsung Exynos 990, but sadly, there aren't any devices with the new MediaTek and Samsung SoCs. Once we get our hands on real devices with the Snapdragon 865, we'll be testing the real-world performance outside of benchmarks, too.

Qualcomm Snapdragon 865, Snapdragon 855, Snapdragon 845, and Kirin 990 Specifications

Quick Overview of Each Benchmark

Benchmark explainer by Mario Serrafero

• AnTuTu: This is a holistic benchmark. AnTuTu tests the CPU, GPU, and memory performance, while including both abstract tests and, as of late, relatable user experience simulations (for example, the subtest which involves scrolling through a ListView). The final score is weighted according to the designer’s considerations.
• GeekBench: A CPU-centric test that uses several computational workloads including encryption, compression (text and images), rendering, physics simulations, computer vision, ray tracing, speech recognition, and convolutional neural network inference on images. The score breakdown gives specific metrics. The final score is weighted according to the designer’s considerations, placing a large emphasis on integer performance (65%), then float performance (30%) and finally crypto (5%).
• GFXBench: Aims to simulate video game graphics rendering using the latest APIs. Lots of onscreen effects and high-quality textures. Newer tests use Vulkan while legacy tests use OpenGL ES 3.1. The outputs are frames during test and frames per second (the other number divided by the test length, essentially), instead of a weighted score.

  GFXBench Subscore Explanations. Click to expand.

  • Aztec Ruins: These tests are the most computationally heavy ones offered by GFXBench. Currently, top mobile chipsets cannot sustain 30 frames per second. Specifically, the test offers really high polygon count geometry, hardware tessellation, high-resolution textures, global illumination and plenty of shadow mapping, copious particle effects, as well as bloom and depth of field effects. Most of these techniques will stress the shader compute capabilities of the processor.
  • Manhattan ES 3.0/3.1: This test remains relevant given that modern games have already arrived at its proposed graphical fidelity and implement the same kinds of techniques. It features complex geometry employing multiple render targets, reflections (cubic maps), mesh rendering, many deferred lighting sources, as well as bloom and depth of field in a post-processing pass.

  read more
• Speedometer, Jetstream: Javascript, core language features and performance on various operations; Javascript math, crypto, and search algorithm performance.
• 3DMark (Sling Shot Extreme OpenGL ES 3.1/Vulkan): The test runs on a mobile-optimized rendering engine using OpenGL ES 3.1 and Vulkan (on Android) or Metal (on iOS). It comes with two subscores, each in turn featuring multiple subscores, all of which ultimately use frames per second as their metric across multiple testing scenarios. This benchmark will test the full range of API features, including transform feedback, multiple render targets and instanced rendering, uniform buffers, and features such as particle illumination, volumetric lighting, deferred lighting, depth of field and bloom in post-processing, all using compute shaders. Offscreen tests use a fixed time step between frames, and rule out any impact caused by vertical sync, display resolution scaling and related OS parameters. The final score is weighted according to the designer’s considerations.
• PCMark 2.0:  Tests the device as a complete unit. It simulates everyday use cases that can implement abstract algorithms and a lot of arithmetic; the difference is that these are dispatched within an application environment, with a particular practical purpose, and handled by API calls and Android libraries common to multiple applications. The test will output a variety of scores corresponding to the various subtests, which will be detailed below; the composite, Work 2.0 score is simply the geometric mean of all of these scores, meaning all tests are weighted equally.

  PCMark 2.0 Subscore Explanations. Click to expand.

  • Web browsing 2.0 simulates browsing social media: rendering the web page, searching for the content, re-rendering the page as new images are added, and so on. This subtest uses the native Android WebView to render (WebKit) and interact with the content, which is locally stored -- this means you can run it offline, but it does not simulate web browsing fully as it rules out internet connection factors (latency, network speed). It is specifically tracking frame rates and completion time across seven tasks, with their score being a multiple of their geometric mean.
  • Video Editing simulates video editing performance: applying effects to a video using OpenGL ES 2.0 fragment shaders, decoding video frames (sent to an Android GLSurfaceView), and rendering/encoding the video in H.264/MPEG-4AVC at several frame rates and resolutions up to 4K. It is specifically tracking frame rates on the UI, except for a final test tracking the completion time of a video editing pipeline.
  • Writing simulates general document and text editing work: adding or editing texts and images within a document, copying and pasting text, and so on. It uses the native Android EditText view as well as PdfRenderer and PdfDocument APIs. It will open compressed documents, move text bodies, insert images in the document, then save them as a PDF, to then encrypt and decrypt them (AES). It specifically tracks task completion times for the processes of opening and saving files, adding images and moving text bodies, encrypt/decrypt the file, and render the PDF pages on ImageViews.
  • Photo Editing simulates photo-editing performance: opening images, applying different effects via filters (grains, blurs, embossing, sharpening and so on) and saving the image. It uses 4MP JPEG source images and manipulates them in bitmap format using the android.media.effect API, android.renderscript API’s RenderScript Intrinsics, android-jhlabs, and the native android.graphics API for drawing the process on the screen. This is an extremely comprehensive test in that it will be impacted by storage access, CPU performance, GPU performance, and it is dependent on many different Android APIs.  The test specifically measures memory and storage access times, encoding and decoding times, task completion times. The various filters and effects come from different APIs.
  • Data manipulation simulates database management operations: parsing and validating data from files, interacting with charts, and so on. It will open (date, value) tuples from CSV, XML, JSON files and then render animated charts with the MPAndroidChart library. It specifically tracks data parsing times as well as draws per second of each chart animation (similar to frame rate, but specific to the updating chart).

  read more

Source links for each benchmark can be found at the end of the article.

Test Devices

*Performance mode on the Snapdragon 865 QRD makes workloads appear 20% "heavier" to the scheduler. This means that a CPU that is loaded 80% will appear 100% loaded to the scheduler, ramping up clocks faster and migrating tasks from the little to the big cores faster. However, CPU clock speeds are NOT boosted.

Benchmark Results

Main Scores

Subscores

Main Scores Comparison

Subscores Comparison

Concluding Highlights

Analysis by Mario Serrafero:

• For AnTuTu’s final score, we observe a large 33% bump over the 855+ and a massive improvement of around 45% over the 855. The CPU subtests showcase massive improvements, with uplifts in each subscore ranging from 15% to 97%. These results are surprising given that Qualcomm posted a respectable 25% CPU performance uplift over the Snapdragon 855, yet we see all CPU subscores go up by over 40%, and even 70%. The GPU side of the subscores, however, sees a much more restrained increase of around 13% on average, compared to the 855+, or 24% to 56% compared to our 855 scores from the Google Pixel 4.
• The popular PCMark 2.0 saw a massive jump of almost 40% in its “Work 2.0” final score, compared to the 855+. Looking at the subscores, it seems that most of the improvement lies in the Photo Editing 2.0 subtest, which nearly doubles in score, followed by a Web Browsing score improvement of around 80%. The final score is simply the average between all subscores, so these massive bumps end up being balancing out the more conservative figures of the other subscores, which remain constant or rise by less than 25%.
• Geekbench 5 subscores gave us a decent look into where the resulting ~20% increase in Single-core and Multi-core scores comes from. The crypto tests (which are weighted the least in calculating the final scores) had a performance increment of 36% and 44% (single and multi, respectively) compared to our 855+ results, whereas integer and floating-point performance only rose by about 19% to 25%, perfectly in-line with Qualcomm’s figures. The gap is much larger if we compare the 865 to our 855 results from the Pixel 4, as crypto goes up by 66% while integer and floating-point improvements sit over 50% for single-core tests and over 35% for multi-core tests. Given the 865 features the same clock speeds as the 855, we see a bump in integer and floating score performance per MHz.
• 3DMark scores also fall more-or-less in line with the expected 20% faster graphics rendering that Qualcomm boasted at the Snapdragon tech summit. The graphics and physics scores saw an increase of 15% and 11% (respectively) over the 855+ for the OpenGL ES 3.1 test, and 25% and 22% for the Vulkan test. This suggests the 865 is a healthy upgrade for gamers.
• GFXBench only saw a performance boost of 5% to 15% over the 855+, though when comparing it against the regular 855 those numbers jump above the 20% year-on-year increments posted by the company.

Recommended Reading

• Qualcomm announces the Snapdragon 865 with support for 5G, 200MP cameras, and 144Hz displays
• Huawei unveils the Kirin 990 with integrated 5G for the Mate 30
• MediaTek announces the Dimensity 1000, a 7nm high-end SoC with integrated 5G
• Samsung announces the 7nm Exynos 990 SoC and the 5G Exynos Modem 5123
• How Qualcomm is Improving Performance, Gaming, and AI on the Snapdragon 855
• Qualcomm unveils the Snapdragon 855 Plus with an overclocked CPU and GPU
• Qualcomm Snapdragon 855 Benchmarks: Comparing the CPU, GPU, and AI Performance with the Kirin 980 and Snapdragon 845
• Qualcomm Snapdragon 845 Benchmarks and Comparison: As Powerful as Promised, for Better or Worse

Benchmark Sources

CPU, GPU, and Memory

CPU and Memory

System

GPU

Storage

Browser

Speedometer 2.0 ||| JetStream 1.1

Thanks to TK Bay for the featured image. Thanks to Max Weinbach for providing the Kirin 990 results from his Huawei Mate 30 Pro.