Qualcomm Snapdragon 865 Benchmarks: Evaluating CPU and GPU Efficiency with the Kirin 990, Snapdragon 855, and Snapdragon 845

Virtually two weeks up to now, Qualcomm invited tech journalists to Maui for the 2019 Snapdragon Tech Summit. On the event, the company unveiled its latest high-end SoC for mobile devices: the Qualcomm Snapdragon 865 cellular platform. Qualcomm says the model new Snapdragon 865 boasts a 25% CPU effectivity enhance and a 20% GPU effectivity enhance over the sooner period Snapdragon 855. Moreover, the model new SoC helps LPDDR5 memory and is manufactured on a newer 7nm course of. Qualcomm’s latest silicon will make its strategy to 2020 flagships similar to the Xiaomi Mi 10, OPPO Discover X2, and loads of totally different high-end smartphones.

Nevertheless merely how quite a bit sooner is it than the sooner generations? We benchmarked Qualcomm’s Snapdragon 865 reference system on the event to hunt out out. We pit the model new SoC in opposition to the Snapdragon 855+, the Snapdragon 855, the Snapdragon 845, and the Kirin 990 from Huawei’s HiSilicon. We might have cherished to verify the Snapdragon 865 in opposition to the MediaTek Dimensity 1000 or Samsung Exynos 990, nonetheless sadly, there aren’t any devices with the model new MediaTek and Samsung SoCs. As quickly as we get our palms on precise devices with the Snapdragon 865, we’ll be testing the real-world effectivity exterior of benchmarks, too.

---

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

	Qualcomm Snapdragon 865	Qualcomm Snapdragon 855+	Qualcomm Snapdragon 855	Qualcomm Snapdragon 845	HiSilicon Kirin 990 (4G)
CPU	1 Kryo 585 ‘Prime’ (ARM Cortex-A77-based), as a lot as 2.84GHz Three Kryo 585 ‘Performance’ (ARM Cortex-A77-based), as a lot as 2.4GHz 4 Kryo 385 ‘Efficiency’ (ARM Cortex-A55-based), as a lot as 1.8GHz 25% Effectivity enchancment over the sooner period	1 Kryo 485 ‘Prime’ (ARM Cortex-A76-based), as a lot as 2.96GHz Three Kryo 485 ‘Performance’ (ARM Cortex-A76-based), as a lot as 2.42GHz 4 Kryo 385 ‘Efficiency’ (ARM Cortex-A55-based), as a lot as 1.8GHz	1 Kryo 485 ‘Prime’ (ARM Cortex-A76-based), as a lot as 2.84GHz Three Kryo 485 ‘Performance’ (ARM Cortex-A76-based), as a lot as 2.42GHz 4 Kryo 385 ‘Efficiency’ (ARM Cortex-A55-based), as a lot as 1.8GHz 45% Effectivity enchancment over the sooner period	4 Kryo 385 ‘Performance’ (ARM Cortex-A75-based), as a lot as 2.8GHz 4 Kryo 385 ‘Efficiency’ (ARM Cortex-A55-based), as a lot as 1.8GHz 25% Effectivity enchancment over the sooner period	2 ARM Cortex-A76, as a lot as 2.86GHz 2 ARM Cortex-A76, as a lot as 2.09GHz 4 ARM Cortex-A55, as a lot as 1.86GHz
GPU	Adreno 650 20% Effectivity enchancment over the sooner period	Adreno 640 (15% overclocked)	Adreno 640 20% Effectivity enchancment over the sooner period	Adreno 630 25% Effectivity enchancment over the sooner period	Mali-G76MP16
Memory	4x 16bit, 2133MHz LPDDR4X 4x 16bit, 2750MHz LPDDR5	4x 16bit, 2133MHz LPDDR4X	4x 16bit, 2133MHz LPDDR4X	4x 16-bit, 1866MHz LPDDR4X	4x 16-bit, LPDDR4X-4266
Manufacturing Course of	7nm (TSMC N7P)	7nm (TSMC)	7nm (TSMC)	10nm LPP (Samsung)	7nm (TSMC)

---

Quick Overview of Each Benchmark

Benchmark explainer by Mario Serrafero

• AnTuTu: This generally is a holistic benchmark. AnTuTu assessments the CPU, GPU, and memory effectivity, whereas along with every abstract assessments and, as of late, relatable particular person experience simulations (as an illustration, the subtest which incorporates scrolling by means of a ListView). The last word ranking is weighted consistent with the designer’s points.
• GeekBench: A CPU-centric check out that makes use of quite a few computational workloads along with encryption, compression (textual content material and footage), rendering, physics simulations, laptop computer imaginative and prescient, ray tracing, speech recognition, and convolutional neural neighborhood inference on images. The ranking breakdown offers specific metrics. The last word ranking is weighted consistent with the designer’s points, inserting a giant emphasis on integer effectivity (65%), then float effectivity (30%) and finally crypto (5%).
• GFXBench: Targets to simulate on-line sport graphics rendering using the latest APIs. Plenty of onscreen outcomes and high-quality textures. Newer assessments use Vulkan whereas legacy assessments use OpenGL ES 3.1. The outputs are frames all through check out and frames per second (the alternative amount divided by the check out measurement, principally), as a substitute of a weighted ranking.

  GFXBench Subscore Explanations. Click on on to develop.

  • Aztec Ruins: These assessments are in all probability essentially the most computationally heavy ones provided by GFXBench. At current, prime mobile chipsets can’t preserve 30 frames per second. Notably, the check out offers really extreme polygon rely geometry, {{hardware}} tessellation, high-resolution textures, worldwide illumination and a great deal of shadow mapping, copious particle outcomes, along with bloom and depth of space outcomes. Most of these strategies will stress the shader compute capabilities of the processor.
  • Manhattan ES 3.0/3.1: This check out stays associated given that fashionable video video games have already arrived at its proposed graphical fidelity and implement the an identical types of strategies. It choices superior geometry utilizing quite a few render targets, reflections (cubic maps), mesh rendering, many deferred lighting sources, along with bloom and depth of space in a post-processing transfer.
• Speedometer, Jetstream: Javascript, core language choices and effectivity on various operations; Javascript math, crypto, and search algorithm effectivity.
• 3DMark (Sling Shot Extreme OpenGL ES 3.1/Vulkan): The check out 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 flip that features quite a few subscores, all of which lastly use frames per second as their metric all through quite a few testing conditions. This benchmark will check out the entire differ of API choices, along with transform solutions, quite a few render targets and instanced rendering, uniform buffers, and choices resembling particle illumination, volumetric lighting, deferred lighting, depth of space and bloom in post-processing, all using compute shaders. Offscreen assessments use a tough and quick time step between frames, and rule out any have an effect on introduced on by vertical sync, present choice scaling and related OS parameters. The last word ranking is weighted consistent with the designer’s points.
• PCMark 2.0:  Assessments the system as an entire unit. It simulates frequently use circumstances that will implement abstract algorithms and various arithmetic; the excellence is that these are dispatched inside an utility ambiance, with a particular wise operate, and handled by API calls and Android libraries widespread to quite a few functions. The check out will output a variety of scores equal to the various subtests, which shall be detailed beneath; the composite, Work 2.Zero ranking is simply the geometric suggest of all of these scores, meaning all assessments are weighted equally.

  PCMark 2.Zero Subscore Explanations. Click on on to develop.

  • Internet looking 2.0 simulates looking social media: rendering the online net web page, looking for the content material materials, re-rendering the net web page as new images are added, and so forth. This subtest makes use of the native Android WebView to render (WebKit) and work along with the content material materials, which is regionally saved — this suggests you might run it offline, nonetheless it would not simulate web looking completely as a result of it pointers out net connection components (latency, neighborhood velocity). It is significantly monitoring physique costs and completion time all through seven duties, with their ranking being a quite a few of their geometric suggest.
  • Video Modifying simulates video enhancing effectivity: making use of outcomes to a video using OpenGL ES 2.Zero fragment shaders, decoding video frames (despatched to an Android GLSurfaceView), and rendering/encoding the video in H.264/MPEG-4AVC at quite a few physique costs and resolutions as a lot as 4K. It is significantly monitoring physique costs on the UI, other than a final check out monitoring the completion time of a video enhancing pipeline.
  • Writing simulates regular doc and textual content material enhancing work: together with or enhancing texts and footage inside a doc, copying and pasting textual content material, and so forth. It makes use of the native Android EditText view along with PdfRenderer and PdfDocument APIs. It ought to open compressed paperwork, switch textual content material our our bodies, insert images inside the doc, then save them as a PDF, to then encrypt and decrypt them (AES). It significantly tracks course of completion events for the processes of opening and saving recordsdata, together with images and shifting textual content material our our bodies, encrypt/decrypt the file, and render the PDF pages on ImageViews.
  • Image Modifying simulates photo-editing effectivity: opening images, making use of completely totally different outcomes by means of filters (grains, blurs, embossing, sharpening and so forth) and saving the image. It makes use of 4MP JPEG provide images and manipulates them in bitmap format using the android.media.impression API, android.renderscript API’s RenderScript Intrinsics, android-jhlabs, and the native android.graphics API for drawing the tactic on the show display. That may be a significantly full check out in that it’s going to be impacted by storage entry, CPU effectivity, GPU effectivity, and it is relying on many different Android APIs.  The check out significantly measures memory and storage entry events, encoding and decoding events, course of completion events. The various filters and outcomes come from completely totally different APIs.
  • Data manipulation simulates database administration operations: parsing and validating data from recordsdata, interacting with charts, and so forth. It ought to open (date, price) tuples from CSV, XML, JSON recordsdata after which render animated charts with the MPAndroidChart library. It significantly tracks data parsing events along with attracts per second of each chart animation (similar to physique value, nonetheless specific to the updating chart).

Provide hyperlinks for each benchmark may very well be found on the end of the article.

---

Check out Models

	Qualcomm Snapdragon 865	Qualcomm Snapdragon 855+	Qualcomm Snapdragon 855	Qualcomm Snapdragon 845	HiSilicon Kirin 990
Gadget Establish	Qualcomm Reference Gadget (QRD)	ASUS ROG Telephone II	Google Pixel 4	Google Pixel 3 XL	Huawei Mate 30 Skilled
Software program program	Android 10 (Qualcomm custom-made AOSP software program program)	Android 9 (ZenUI 6.Zero OEM software program program with October 2019 security patch)	Android 10 (Google Pixel OEM software program program with December 2019 security patch)	Android 10 (Google Pixel OEM software program program with December 2019 security patch)	Android 10 (EMUI 10.Zero OEM software program program with October 2019 security patch)
Present	2880×1440 @ 60Hz	2340×1080 @ 60Hz	2280×1080 @ 60Hz	2960×1440 @ 60Hz	2400×1176 @ 60Hz
Memory	12GB LPDDR5	8GB LPDDR4X	6GB LPDDR4X	4GB LPDDR4X	8GB LPDDR4X
Storage	128GB UFS 3.0	128GB UFS 3.0	64GB UFS 2.1	64GB UFS 2.1	256GB UFS 3.0
Effectivity Mode	Certain*	No	No	No	No

*Effectivity mode on the Snapdragon 865 QRD makes workloads appear 20% “heavier” to the scheduler. Due to this a CPU that is loaded 80% will appear 100% loaded to the scheduler, ramping up clocks sooner and migrating duties from the little to the large cores sooner. However, CPU clock speeds are NOT boosted.

---

Benchmark Outcomes

Most vital Scores

Benchmark	Mannequin	Qualcomm Snapdragon 865	Qualcomm Snapdragon 855+	Qualcomm Snapdragon 855	Qualcomm Snapdragon 845	HiSilicon Kirin 990
AnTuTu	8.0.4	565,384	425,963	386,499	278,647	389,505
Geekbench single-core	5.0.2	929	760	600	521	750
Geekbench multi-core	5.0.2	3,450	2,840	2,499	2,125	2,887
GFXBench ES 3.0 1080 Manhattan offscreen	5.00	126	110	92	82	104
GFXBench ES 3.1 1080 Carchase offscreen	5.00	50	48	40	35	38
GFXBench ES 3.1 1080 Manhattan offscreen	5.00	88	78	67	61	67
GFXBench ES 2.0 1080 T-Rex offscreen	5.00	205	185	164	152	105
GFXBench 1440p Aztec Ruins Vulkan (Extreme Tier) Offscreen IFH	5.00	20	19	16	14	16
GFXBench 1440p Aztec Ruins OpenGL (Extreme Tier) Offscreen IFH	5.00	20	18	16	14	18
Speedometer	2.00	80	36	53	49	65.4
JetStream – Geometric suggest	1.10	123	116	98	85	95.8
PCMark – Work 2.0	2.0.3716	12,626	9,068	9,311	8,988	8,667
Androbench Sequential Be taught (MB/s)	5.0.1	1,459	1,398	873	659	1,451.09
Androbench Sequential Write (MB/s)	5.0.1	225	217	189	231	443.66
Androbench Random Be taught (IOPS)	5.0.1	50,378	41,315	37,600	32,376	53,114.78
Androbench Random Write (IOPS)	5.0.1	48,410	35,422	41,340	37,417	55,972.18
Androbench Random Be taught (MB/s)	5.0.1	195	161	147	126	207.47
Androbench Random Write (MB/s)	5.0.1	189	138	161	146	218.64
Androbench SQLite Insert	5.0.1	3,705	3,187	3,207	2,627	4,968.81
Androbench SQLite Exchange	5.0.1	4,014	3,931	3,996	3,333	6,090.65
Androbench SQLite Delete	5.0.1	5,037	4,964	4,558	4,081	7,664.88
3DMark Sling Shot Extreme Open GL ES 3.1 Complete Ranking	2.0.4646	7,008	6,201	5,174	3,431	5,677
3DMark Sling Shot Extreme Vulkan Complete Ranking	2.0.4646	6,449	5,339	4,339	3,273	4,303

Subscores

Most vital Scores Comparability

Subscore	Versus Snapdragon 865	Versus Snapdragon 855+	Versus Snapdragon 855	Versus Snapdragon 845	Versus Kirin 990
AnTuTu	1x	1.33x	1.46x	2.03x	1.45x
Geekbench single-core	1x	1.22x	1.55x	1.78x	1.24x
Geekbench multi-core	1x	1.21x	1.38x	1.62x	1.2x
GFXBench ES 3.0 1080 Manhattan offscreen	1x	1.15x	1.37x	1.54x	1.21x
GFXBench ES 3.1 1080 Carchase offscreen	1x	1.04x	1.25x	1.43x	1.32x
GFXBench ES 3.1 1080 Manhattan offscreen	1x	1.13x	1.31x	1.44x	1.31x
GFXBench ES 2.0 1080 T-Rex offscreen	1x	1.11x	1.25x	1.35x	1.95x
GFXBench 1440p Aztec Ruins Vulkan (Extreme Tier) Offscreen IFH	1x	1.05x	1.25x	1.43x	1.25x
GFXBench 1440p Aztec Ruins OpenGL (Extreme Tier) Offscreen IFH	1x	1.11x	1.25x	1.43x	1.11x
Speedometer	1x	2.22x	1.51x	1.63x	1.22x
JetStream – Geometric suggest	1x	1.06x	1.26x	1.45x	1.28x
PCMark – Work 2.0	1x	1.39x	1.36x	1.4x	1.46x
Androbench Sequential Be taught (MB/s)	1x	1.04x	1.67x	2.21x	1.01x
Androbench Sequential Write (MB/s)	1x	1.04x	1.19x	0.97x	0.51x
Androbench Random Be taught (IOPS)	1x	1.22x	1.34x	1.56x	0.95x
Androbench Random Write (IOPS)	1x	1.37x	1.17x	1.29x	0.86x
Androbench Random Be taught (MB/s)	1x	1.21x	1.33x	1.55x	0.94x
Androbench Random Write (MB/s)	1x	1.37x	1.17x	1.29x	0.86x
Androbench SQLite Insert	1x	1.16x	1.16x	1.41x	0.75x
Androbench SQLite Exchange	1x	1.02x	1x	1.2x	0.66x
Androbench SQLite Delete	1x	1.01x	1.11x	1.23x	0.66x
3DMark Sling Shot Extreme Open GL ES 3.1 Complete Ranking	1x	1.13x	1.35x	2.04x	1.23x
3DMark Sling Shot Extreme Vulkan Complete Ranking	1x	1.21x	1.49x	1.97x	1.50x

Subscores Comparability

---

Concluding Highlights

Analysis by Mario Serrafero:

• For AnTuTu’s final ranking, we observe a giant 33% bump over the 855+ and an unlimited enchancment of spherical 45% over the 855. The CPU subtests showcase enormous enhancements, with uplifts in each subscore ranging from 15% to 97%. These outcomes are gorgeous given that Qualcomm posted an excellent 25% CPU effectivity uplift over the Snapdragon 855, however we see all CPU subscores go up by over 40%, and even 70%. The GPU side of the subscores, nonetheless, sees a far more restrained enhance of spherical 13% on frequent, as compared with the 855+, or 24% to 56% as compared with our 855 scores from the Google Pixel 4.
• The favored PCMark 2.0 seen an unlimited bounce of nearly 40% in its “Work 2.0” final ranking, as compared with the 855+. Making an attempt on the subscores, it seems that evidently numerous the enchancment lies inside the Image Modifying 2.Zero subtest, which virtually doubles in ranking, adopted by a Internet Looking out ranking enchancment of spherical 80%. The last word ranking is simply the standard between all subscores, so these enormous bumps end up being balancing out the additional conservative figures of the alternative subscores, which keep fastened or rise by decrease than 25%.
• Geekbench 5 subscores gave us an trustworthy look into the place the following ~20% enhance in Single-core and Multi-core scores comes from. The crypto assessments (which can be weighted the least in calculating the final word scores) had a effectivity increment of 36% and 44% (single and multi, respectively) as compared with our 855+ outcomes, whereas integer and floating-point effectivity solely rose by about 19% to 25%, utterly in-line with Qualcomm’s figures. The opening is way greater if we look at the 865 to our 855 outcomes from the Pixel 4, as crypto goes up by 66% whereas integer and floating-point enhancements sit over 50% for single-core assessments and over 35% for multi-core assessments. Given the 865 choices the an identical clock speeds as a result of the 855, we see a bump in integer and floating ranking effectivity per MHz.
• 3DMark scores moreover fall more-or-less based on the anticipated 20% sooner graphics rendering that Qualcomm boasted on the Snapdragon tech summit. The graphics and physics scores seen an increase of 15% and 11% (respectively) over the 855+ for the OpenGL ES 3.1 check out, and 25% and 22% for the Vulkan check out. This implies the 865 is a healthful enhance for gamers.
• GFXBench solely seen a effectivity enhance of 5% to fifteen% over the 855+, though when evaluating it in opposition to the frequent 855 these numbers bounce above the 20% year-on-year increments posted by the company.

---