Qualcomm Snapdragon 845 Hands On: Benchmarks and First Impressions


The Qualcomm Snapdragon 845 was officially announced last December, though its reveal at the annual Snapdragon Tech Summit left us with almost as many questions as answers. While we managed to get a surface-level description of its architecture and capabilities, so far, we’ve had to rely on the company’s internal data — namely, its quoted year-on-year percentage increments — to estimate the new platform’s performance. Now, we’ve got benchmark scores.

This week, a group of journalists, analysts, and YouTube personalities were invited to Qualcomm’s 5G Day event, where the company released more information about its connectivity efforts and the future of the mobile internet. Afterward, some of us got to stay for a benchmarking session with a reference device sporting the Snapdragon 845 and other high-end components. While we only had around two to three hours of hands-on time with the device — and despite the fact that the reference device was built for the sole purpose of testing (and now showcasing) the platform — we managed to gain insight on what to expect from upcoming flagship devices that’ll feature the Snapdragon 845.

Before we show you some of the results we collected, here’s a quick refresher on the Snapdragon 845, including what’s changed and what’s new in terms of CPU and GPU design and implementation.


A Bit of Background

The Snapdragon series’s chip architecture, which has historically featured a mix of custom and semi-custom cores based on ARM designs, has improved dramatically over the past decade. Qualcomm’s Scorpion CPU core was followed by its custom Krait CPU core, starting with the Snapdragon S4 in 2012. In 2015, Qualcomm moved to a combo of 64-bit stock ARM Cortex-A57 and Cortex-A53 cores with the Snapdragon 810 and 808, retiring Krait in the process. But only a year later, Qualcomm was back in the custom CPU core game with the Snapdragon 820. It marked the debut of Kryo (featured in comparisons below), which placed a heavy emphasis on floating point IPC (Instructions Per Clock) in single-threaded performance.

Kryo’s CPU performance and power efficiency improved on Qualcomm’s rather underwhelming implementation of the ARM Cortex-A57 in the Snapdragon 808 and 810, but benchmarks showed that it couldn’t match ARM’s 2016 core, the Cortex-A72, in terms of integer IPC. That said, it was a redeeming release for Qualcomm; its predecessor had tarnished the company’s reputation among some reviewers who, in many cases, couldn’t ignore the heat and throttling problems seen on many Snapdragon 810 devices, particularly earlier models like the HTC One M9 and LG G Flex 2.

With the Snapdragon 835, Qualcomm changed things up again with “semi-custom” CPU cores that took advantage of the “Built on ARM Cortex Technology” license. The Snapdragon 835 features Kryo 280 “performance” cores based on ARM’s A73 design that are faster than the company’s last-gen fully-custom predecessors in terms of integer instructions per clock (IPC), but regress when it comes to floating-point math (FPM). Still, the Snapdragon 835 remains one of the fastest system-on-chips in the Android market, and it’s a substantial leap forward from a technological standpoint, bringing better power efficiency and thermal stability as well as advancements in peripheral components.

Snapdragon 845 Improvements Overview

Specs Qualcomm Snapdragon 845 Qualcomm Snapdragon 835
Chipset 845 (10nm LPP) 835 (10nm LPE)
CPU 4x 2.8GHz Kryo 385 (A75 “performance”), 4x 1.8GHz Kryo 385 (A55 “efficiency”) 4x 2.45GHz Kryo 280 (A73 big), 4x 1.9GHz Kryo 280 (A53 LITTLE)
GPU Adreno 630 GPU Adreno 540 GPU
Memory 4x 1866MHz 32-bit LPDDR4X 4x 1866MHz 32-bit LPDDR4X
ISP/Camera Dual 14-bit Spectra 280 ISP 32MP Dual 14-bit Spectra 180 ISP 32MP
Modem Snapdragon X20 LTE (Cat 18 downlink, Cat 13 uplink) Snapdragon X16 LTE (Cat 16 downlink, Cat 13 uplink)

As you might have noticed, the Snapdragon 845 is the first Qualcomm chip in multiple generations that hasn’t been architecturally overhauled in a switch from custom to semi-custom cores, or vice-versa. It re-employs the “Built on ARM” license, following in the footsteps of last year’s Snapdragon 835. This marks the first time in years we’ve seen Qualcomm flagships stick with custom or semi-custom core design two years in a row, and it’s not unjustified. The Snapdragon 845 features eight Kryo 385 CPU cores, and while their name suggests homogeneity, it actually consists of four Cortex-A75 performance cores and four Cortex-A55 efficiency cores. The jump to newer cores would by itself suggest a healthy boost in performance, as would the adoption of Samsung’s 2nd-generation 10nm LPP (low power plus) FinFET process on which the chip is built. Those updates and other improvements contribute to the quoted 30% performance uplift over last year’s 835, and the 25% to 30% overall improvement in power-efficiency.

Snapdragon 845 System-on-chip (Source: Qualcomm)

Source: ARM

The Kryo 385’s performance (“Gold”) cores are clocked up to 2.8GHz, up from the Kryo 280’s 2.4GHz. The A75 design improves upon previous years’ A72 and A73 in terms of performance, while moving to the ARMv8.2 architecture, which brings an enhanced memory model, scalable vector extensions (SVE), and other enhancements. The cores also add features like support for ARM’s DynamIQ, ARM’s improved standard for heterogeneous computing.

The A72 and A73 focused greatly on improving thermal stability and power-efficiency, and the A75 carries over those benefits (for example, by keeping the A73’s branch predictor with minimal tuning) while exhibiting a concerted improvement in performance.

The A75 has a 22% uplift over the Cortex-A73 on the same process node and at the same clockspeed. It sees over 20% better integer core performance, and 33% higher floating-point and NEON performance (with the addition of support for FP16 half-precision processing), and improvements in machine learning performance through the inclusion of an INT8 dot product instruction for 8-bit neural-network algorithms (although you would still likely want to execute machine learning workloads on the Snapdragon 845’s Adreno 630 GPU or compute DSP). When the A75 was originally unveiled and detailed, ARM suggested that we could expect a 34% increase in Geekbench performance compared to the Cortex-A73, which saw low double-digit percentage improvements over A72 at most. In just a few more paragraphs, we’ll see how that translates to the Snapdragon 845.

Advantages of heterogeneous computing. (Source: Qualcomm)

DynamIQ is also a promising advancement, building upon big.LITTLE to make the most out of the A75+A55 combo found on the Snapdragon 845. DynamIQ governs the grouping of CPU clusters and their inter-communication for heterogeneous computing. It supports up to eight CPUs per cluster, with up to eight voltage/frequency domains per CPU cluster — Snapdragon 845 has a familiar two-cluster setup, with three clock and voltage domains. The bridge between clusters is carried out by a DynamIQ shared unit, or DSU, which can host an optional shared L3 cache (with the A75/A55 now having private L2 caches instead), and the Snapdragon 845 takes full advantage of it. DynamIQ also enables finer-grain CPU clock speed control, which the 845 will handily utilize.

While we are on the subject of shared caches, the Snapdragon 845 in particular also offers a separate 3MB system cache for all SoC blocks, which Qualcomm claims can help reduce access transactions by up to 75%, in turn yielding some performance and power saving improvements.

Source: ARM

A55 vs. A53 (Source: ARM)

The Kryo 385’s (“Silver”) cluster features “efficiency” cores based on ARM’s Cortex-A55 and clocked at 1.8GHz. Qualcomm claims that the resulting performance uplift is around 15%, and the company also noted that the cores play a key part in the heterogeneous compute platform’s overall power efficiency. Indeed, we’ve seen great results with previous-generation efficiency cores in Qualcomm’s flagship chipsets, but also in the mid-range (The Snapdragon 625, which featured A53 cores exclusively and had legendary endurance, is a prime example). The A55 sees the expected improvements such as the aforementioned ARMv8.2 architecture extensions, dedicated machine learning instructions, and private L2 cache (up to 256KB), and also a redesigned micro-architecture that promises a 18% performance improvement performance at 15% better power-efficiency (we’ll have to see how Qualcomm decided to tune those knobs, but it’ll likely be in favor of endurance).

That 18% performance reference increase is reflected across 18% better integer performance, 20% higher floating point performance, 40% higher performance in NEON SIMD and 15% faster JavaScript, alongside a massive boost of up to 200% to memory-bound workloads according to ARM. The reduced cache latency and performance optimizations make it an overall-better version of the power-efficient core behind last year’s notable endurance kings, and with the 845 featuring a slightly-lower frequency in the efficiency cluster (by 100MHz compared to the 835), we expect this A55 arrangement to be a big contributor to battery life savings.

Last but not least, the Snapdragon 845 brings the expected improvements to Qualcomm’s custom GPU line, with the new Adreno 630 promising 30% faster performance while also remaining 30% more power efficient. Unlike with ARM-based CPUs on the 845, it’s been a challenge to uncover specifics about what’s new and improved beyond performance numbers — we know that it has twice as many compute cores as the previous-generation Adreno GPU, for example… but not much else.

We’ve been treated to larger year-on-year proportional GPU improvements in the past, but it’s worth noting that Qualcomm’s GPUs in particular stand above competitors in the Android space, something that can’t always be said about its CPU offerings. The Mali-G72 (12 core variant) featured in the HiSilicon 970 and the Mali-G71 (20 core variant) found in the Exynos 8895 began bridging that performance gap, but at the expense of power efficiency. This is important for Qualcomm, given that the company is focusing on heterogeneous computing in a unified platform, and the improvements in power efficiency across the board play a large role in that. It also fits into the company’s focus on virtual reality (it’s no surprise that Snapdragon chipsets are making their way to VR headsets), and on-device machine-learning efforts (its SDKs allow developers to distribute workloads across the CPU, GPU and compute DSP as necessary).

Testing Unit, Methodology & Pitfalls

Qualcomm Snapdragon 845 Reference Design OnePlus 5 (Snapdragon 835) OnePlus 3T (Snapdragon 821)
Android Version Android 8.0 Oreo OxygenOS 5.0.2, Android 8.0 Oreo OxygenOS 5.0.1, Android 8.0 Oreo
Chipset Snapdragon 845 (Octa-core, 10nm, 4x 2.8GHz + 4x 1.8GHz) Qualcomm Snapdragon 835 (Octa-core, 10nm, 4x 2.45GHz + 4x 1.9GHz) Qualcomm Snapdragon 821/MSM8996 Pro (Quad-core, 14nm, 2x 2.4 GHz + 2x 1.6 GHz)
GPU Adreno 630 GPU Adreno 540 GPU Adreno 530 GPU
Display 5.5-inch 2560 x 1440 pixels (538 ppi) 5.5-inch 1920 x 1080 pixels (401 ppi) 5.5-inch 1920 x 1080 pixels (401 ppi)
Storage UFS 2.1 UFS 2.1 UFS 2.0

Once it came time to test the Snapdragon 845, we were taken to a small conference room in Qualcomm’s San Diego headquarters where we had a few hours with the latest hardware from Qualcomm’s Reference Design program. This unit resembled something that could actually be sold at a store, unlike the brute, glossy brick that was the Snapdragon 835 reference model (MDP/S). It sported a 5.5-inch QHD display and powerful components including a modest camera sensor, detailed in the table above this paragraph. Qualcomm has been focusing on developing a more thermally stable platform, and that was evident from the reference design’s performance — the device was impressively thermally stable, maintaining scores within the expected ranges even at higher temperatures.

It was running Android 8.0.0 Oreo without modifications, but the device had USB debugging enabled once we got to it, and root access had seemingly been enabled as well (we couldn’t take advantage of that right then and there). It had been used for benchmarking several times before our session, with scores dating weeks back that were appreciably lower than those we obtained.


A few words on methodology: We only had a few hours with the Snapdragon 845 reference device, and it must be noted that the ROM it was running was far from a production-ready package. We were briefed ahead of time on some testing anomalies that we had to watch out for, so the results we obtained shouldn’t have been impacted by the device’s software. That being said, some tests like PCMark rely on Android API calls and thus might be more susceptible to extraneous behavior introduced by the ROM, and our smoothness tests are also heavily dependent on ROM optimization. We expect some of these numbers to be slightly different than those we’ll report in the future, once we get to test the Snapdragon 845 on actual production units. OEMs will be introducing their own kernel and governor changes, and they will ultimately dictate how the processor performs on their devices (potentially not using the same schedutil CPU scaling governor that the reference device uses). Still, these benchmarks should still give us an informed preview of what to expect.

Because we had a limited amount of time with these devices, and because each of us was only given one unit to test, we couldn’t afford to thoroughly verify that confounders weren’t, in fact, altering the scores. That said, we also have no reason to believe these scores aren’t reliable: we independently disabled the few apps on the device to prevent them from running in the background (and appreciably but minimally affecting score points), and all of our results fell within (or above) Qualcomm’s proposed ranges. One issue we certainly could not avoid was heat, as the time constraints forced us to run most of the benchmark tests sequentially. We did allow the device to cool down after the longer graphics-intensive tests, though, and as we said before, we don’t think that the heat introduced significant throttling (we didn’t observe appreciable changes in the CPU frequency graphs).

We performed every test three times, with the exception of Geekbench (four times) and PCMark (one time). In order to compare the changes across system-on-chip generations, we ran the same benchmarks the same number of times on a OnePlus 3T (6GB) and OnePlus 5 (6GB). Both of these devices have 1080p displays, so we’ve only included off-screen graphics tests in this comparison. However, near the end of the article, you will find a link to all of the data we used for this article, where you’ll also see on-screen 1440p results for the SDM845. Without further ado, here are the numbers!

Benchmark Test Results

First up, we’ll be taking a look at Geekbench 4, one of the better (if not the best) test for assessing CPU performance on Android devices and across platforms. This benchmark has been extremely popular among enthusiasts for many years, and the team behind it has been listening to both users and companies to optimize the accuracy and maximize the usefulness of its tests. Geekbench 4 introduced a new score scale normalized around the Intel Core i7-6600U (which has a baseline score of 4,000), as well as some pauses in-between workloads to minimize the effect of thermal throttling (as a result, it has a longer completion time than Geekbench 3). The 4.1 update also improved multi-core scalability and made changes to the memory latency workload to avoid cache hits on system-on-chips with Cortex-A72 and A73 cores (this is one of the reasons we had to re-test some of our scores for this article, as single-core and multi-core scores saw a slight increase of around 2% and 5% respectively). Geekbench 4 uses tests that implement popular algorithms and workloads homologous to those behind the scenes in many popular applications, so its scores are very insightful. The detailed breakdown will help us assess some of the improvements on Qualcomm’s new chipset.

Geekbench 4
Geekbench 4

Snapdragon 845

With the Snapdragon 845, we see improvements across the board, something that couldn’t be said of last year’s flagship system-on-chip. The single-core score sees an average increase of 25%, while the multi-core score sees a smaller uplift of 24%. Those figures are around the expected improvements of 25% to 30%, and for the most part, we see an increase in each of the sub-scores in Geekbench (see the chart below). Another interesting observation is that both floating point score per MHz and integer score per MHz show improvement relative to the Snapdragon 835. The cores in last year’s Snapdragon 835 saw an increase in integer score per MHz, but a decline in floating point score per MHz compared to the Krait cores in the Snapdragon 821. This time, there’s less compromise (and to be clear, compromise is not what we want here) from one generation to the next in those categories, and the higher clockspeed of the 845 means that this per MHz advantage should translate into the expected performance uplift.

Snapdragon 845

SDM845 Single-Core Performance Improvement SDM835 Single-Core Performance Improvement MSM8996
Single 2453 x1.25 1965 x1.06 1841
Crypto 1547 x1.27 1223 x1.58 776
Integer 2759 x1.33 2074 x1.12 1859
Floating Point 2065 x1.45 1422 x0.84 1696
Memory Score 2570 x.94 2721 x1.19 2285
AES (GB/sec) 1.16 x1.23 942.4 x1.78 529.8
LZMA (MB/sec) 4.14 x1.45 2.86 x1.29 2.22
JPEG (Mpixels/sec) 21.9 x1.32 16.6 x0.75 22
Canny (Mpixels/sec) 32.3 x1.27 25.5 x0.79 32.1
Lua MB/sec) 2.20 x1.25 1.76 x1.24 1.42
Dijkstra (MTW/sec 1.88 x1.08 1.74 x1.20 1.45
SQLite (Krows/sec) 71.8 x1.35 53.3 x1.43 37.2
HTML5 Parse (MB/sec) 12.9 x1.43 8.99 x1.01 8.90
HTML5 DOM (KElements/sec) 2930 x1.31 2230 x2.97 746.6
Histogram (Mpixels/sec) 68.4 x1.31 52.2 x0.92 56.7
PDF Render (Mpixels/sec) 68.6 x1.37 50.1 x0.84 59.5
LLVM (functions/sec) 353.8 x1.35 262.6 x1.58 165.9
Camera (images/sec) 7.82 x1.38 5.68 x0.74 7.70
N-Body Physics (Kpairs/sec) 1440 x1.64 877.8 x0.79 1110
Ray Tracing (Kpixels/sec) 353.5 x1.51 233.4 x0.81 286.7
Rigid Body Physics (FPS) 8683.3 x1.40 6189.4 x1.06 5815.2
HDR (Mpixels/sec) 12 x1.42 8.48 x0.71 12
Gaussian Blur (Mpixels/sec) 33.9 x1.40 24.3 x0.48 51.1
Speech Recognition (Words/sec) 18.7 x1.30 14.4 x1.36 10.6
Face Detection (Ksubwindows/sec) 823.8 x1.62 509.1 x0.76 671.7
Memory Copy (GB/sec) 6.04 x1.22 4.94 x0.77 6.38
Memory Latency (ns) 174.9 x1.40 124.8 x0.53 237
Memory Bandwidth (GB/sec) 15.9 x0.86 18.5 x1.53 12.1


SDM845 Multi-Core Performance Improvements SDM835 Multi-Core Performance Improvements MSM8996
Multi 8437 x1.24 6788 x1.66 4104
Crypto 7025 x1.15 6117 x3.04 2013
Integer 11071 x1.23 8981 x1.84 4879
Floating Point 8288 x1.33 6232 x1.51 4134
Memory Score 3087 x1.05 2937 x1.03 2838
AES (GB/sec) 5.28 x1.14 4.62 x3.12 1.48
LZMA (MB/sec) 15.4 x1.17 13.2 x1.92 6.87
JPEG (Mpixels/sec) 98.4 x1.22 80.9 x1.66 48.7
Canny (Mpixels/sec) 142.2 x1.17 121.5 x1.59 76.6
Lua MB/sec) 8.40 x1.05 8.03 x2.01 4
Dijkstra (MTW/sec 7.14 x1.31 5.47 x1.49 3.66
SQLite (Krows/sec) 309 x1.32 234.4 x2.41 97.4
HTML5 Parse (MB/sec) 58.1 x1.39 41.9 x1.79 23.4
HTML5 DOM (KElements/sec) 7.14 x1.43 5.01 x2.66 1.88
Histogram (Mpixels/sec) 303 x1.18 256.1 x1.72 149
PDF Render (Mpixels/sec) 306.2 x1.21 252.2 x1.99 126.5
LLVM (Kfunctions/sec) 1440 x1.20 1200 x2.46 488.3
Camera (images/sec) 34 x1.28 26.6 x1.58 16.8
N-Body Physics (Mpairs/sec) 6.04 x1.48 4.07 x1.67 2.44
Ray Tracing (Kpixels/sec) 1420 x1.64 1010 x1.64 616.6
Rigid Body Physics (FPS) 39598 x1.38 28718.4 x1.70 16915.3
HDR (Mpixels/sec) 51.3 x1.30 39.6 x1.64 24.2
Gaussian Blur (Mpixels/sec) 142.7 x1.32 108.3 x1.43 75.7
Speech Recognition (Words/sec) 52.2 x1.17 44.6 x1.42 31.4
Face Detection (Ksubwindows/sec) 3.31 x1.40 2.37 x1.25 1.89
Memory Copy (GB/sec) 9.11 x1.29 7.07 x.71 9.96
Memory Latency (ns) 167.8 x1.29 130.1 x0.55 237.2
Memory Bandwidth (GB/sec) 18.6 x1.20 15.5 x0.88 17.6

Overall, Geekbench 4 shows a healthy (if unspectacular) year-on-year improvement. But crucially, the scores aren’t enough to beat Apple’s A11 Bionic system-on-chip, which scores over 4,200 in single-core tests and over 10,100 in multi-core tests. Ever since Apple began running away with chip benchmarks a few years ago, the gap has only grown larger between it and Qualcomm, to the point where latter’s claims of 25% to 30% year-on-year improvements with each Snapdragon revision have become a sign of its inability to topple Apple’s custom silicon in this regard.

Of course, there are some counter-arguments that serve to undermine the comparison. The seemingly insurmountable gap between Qualcomm and Apple’s system-on-chips shrinks when you consider metrics such as performance per square millimeter, for instance, or when you look at the particular goals of each company. Qualcomm intends the Snapdragon 845 to have a performance-to-watt-to-square-millimeter ratio that best serves applications not just on smartphones, but also on virtual reality headsets, connected devices, and Windows computers. Apple designs its chipsets primarily, and almost exclusively, with one device in mind: the iPhone.

Arguments and counter-arguments on that point aside, the performance increment for the Snapdragon 845 is around what we anticipated and what’s been claimed by Qualcomm. Just don’t expect the CPU capabilities of the 845 (and certainly not its Geekbench score) to match Apple’s current and upcoming chipsets.

AnTuTu Benchmark
AnTuTu Benchmark

Snapdragon 845

SDM845 Performance Improvement SDM835 Performance Improvement MSM8996
Overall 265569 x1.24 213994 1.23 173450
CPU 91838 x1.25 73254 1.35 54085
GPU 107322 x1.25 85999 1.24 69286
UX 58498 x1.89 30918 .74 42047
MEM 7910 x.75 10489 1.31 8033

Moving on, we have benchmark results from AnTuTu, an extremely popular and holistic test that gets meaningful revisions on a frequent basis. While AnTuTu is perhaps best known for its prominent one-score test result, it’s the individual sub-score breakdown that best enables us to assess the differences between chips in most instances, and in this case especially.

The UX and memory tests involve components and factors beyond the CPU and GPU we are focusing on, and so deviations from our projected results aren’t entirely unexpected. Even so, the average score increase for the Snapdragon 845 sits comfortably at the expected 25% range, as do the average GPU and CPU scores. The UX test, which simulates real-world application usage (e.g., list scrolling, loading text and image elements, and so on), sees a gigantic boost over our OnePlus 5-specific score, while the memory workload sees a reduction. Given that the final score is a sum of all independent scores, it’s this UX test that disproportionately impacts the final result in the 845’s favor. Because it’s a test acutely influenced by system behavior, we recommend paying less attention to it.

PCMark for Android Benchmark
PCMark for Android Benchmark

Snapdragon 845

SDM845 Performance Improvement SDM835 Performance Improvement MSM8996
Web 2.0 Score 8197 x1.23 6667 x1.14 5828
Web Browsing 6971 x1.10 6321 x1.20 5263
Video Editing 5726 x1.11 5146 x1.13 4542
Writing 8278 x1.25 6604 x1.37 4821
Photo Editing 17196 x1.55 11060 x.90 12273
Data Manipulation 6515 x1.18 5543 x1.17 4752

Another test that both simulates real-world applications and usage scenarios and that’s thoroughly dependent on ROM and kernel/governor tuning is PCMark. We don’t know much about the overall behavior of Qualcomm’s reference design, so we can’t comment on just how homologous to a retail unit the company’s reference design frequency scaling might be. As we’ve seen across reviews, PCMark scores tend to vary from phone to phone, even when said phones feature similar or identical specifications. With all of that said, most tests see a marked double-digit increase on the Snapdragon 845, with the exception of the photo-editing test. (Full disclosure: we could only record one score for this test, as we had some difficulties both installing and running the benchmark on several test units.)

GFXBench GL Benchmark
GFXBench GL Benchmark

3DMark - The Gamer's Benchmark
3DMark - The Gamer's Benchmark

Snapdragon 845

Snapdragon 845

3DMARK SDM845 Performance Improvement SDM835 Performance Improvement >MSM8996
Score 4859 x1.18 4103 1.40 2924
Physics 5444 x1.75 3112 1.55 2010
Graphics 3515 x.78 4513 1.34 3362
G1 31.8 x1.11 28.7 1.24 23
G2 18.9 x1.27 14.9 1.40 10.7
P1 58.7 x1.09 54 1.11 48.8
P2 35.6 x1.05 34.1 1.52 22.4
P3 20.4 x1.20 17 1.78 9.57

Moving on to graphics benchmarks, we took a look at GFXBench’s popular Manhattan (ES 3.1) and Car Chase tests and 3DMark’s Slingshot Unlimited test (ES 3.1). (We didn’t run through Vulkan and haven’t included the on-screen results of graphics tests in this comparison, though you’ll be able to find the on-screen scores in our spreadsheet.) It’s in these tests that we see some of the beefier performance figures brought forth by Qualcomm’s Adreno 630 GPU. Specifically, we see double-digit improvements approaching (and in some cases exceeding) a 50% performance boost on GFXBench’s Manhattan and Car Chase offscreen tests, while 3DMark sees an increase of 18% in overall score. The physics score sees the biggest improvement, with a 75% higher score and variable increases in the three portions of the test.

We also ran the Manhattan ES 3.1 Endurance / Battery Life test on the Snapdragon 845, a 30-minute test that pushes the thermal envelope of whatever device it’s run on (with the Snapdragon 845 in particular, we saw an absurd peak surface temperature of 47°C | 117°F), and despite the unit becoming unbearably hot, the framerate only dropped around 16%, and stabilized higher near the end of the test. This is certainly not bad considering that we normally make sure to start this test at a cool 28°C | 82.4°F, a luxury we couldn’t afford in a (literally) heated benchmarking session.  We’ve provided some graphs comparing the throttling across the 821 and 835, but keep in mind those results were obtained in much more controlled testing environments — I wouldn’t draw strong conclusions from these specific results.

Snapdragon 845

Last but not least on the list of synthetic benchmarks, we have a group of browser tests: Octane, Kraken, Jetstream, and Sunspyder. Fortunately, the Snapdragon 845 showed a year-on-year improvement in final score relative to the Snapdragon 835 on these tests. We’ve included the full score breakdown on the spreadsheet at the bottom of this article, and we suggest you refer to that sheet, given we were able to record a lot more scores for each specific workload. It’s simply infeasible for us to include all of those breakdowns in this article without impacting readability, so we chose to focus on the more popular scores and tests.

We ran a couple of other tests that didn’t produce significant results. Geekbench 4’s RenderScript score showed a massive 100% uplift over the Snapdragon 835, with the Snapdragon 845 achieving a score of 14,353 and the Razer Phone and Exynos S8-based devices scoring in the 8,000 range. A few members of the press at the benchmarking session, including Fudzilla’s Fuad Abazovic inquired about this, and were informed that it might be related to the twofold increase in the number of compute cores in the Snapdragon 845 (we were told that graphics performance, however, is limited by a fixed pipeline, so don’t expect to see such a dramatic improvement in most workloads). We also ran one of our smoothness tests on the Snapdragon 845 just for kicks to see whether the reference device’s Oreo ROM was well-optimized and/or whether the 845 showed measurable advantage in UI performance… pointlessly so, admittedly, because it’s impossible for us to determine whether either, both, or none are true. That said, the Play Store scrolling test (a simple multi-second set of fast swipes through a preloaded “Top Charts” list) showed rather amazing results (graphs above).

Benchmarks Giveth and Benchmarks Taketh Away

We’ve gone through a plethora of benchmarks and have been able to get a glimpse of the Snapdragon 845’s performance. However, there’s still a lot to uncover, and how the system-on-chip ultimately performs will depend on manufacturer implementations. We hope this has been a useful, if imperfect, comparison. We’ll certainly be revisiting the Snapdragon 845 — and its instantiation in 2018 devices — once flagship phones begin rolling out.

With the wealth of benchmark information we’ve unpacked, there are a few key takeaways. Qualcomm’s claims of a 30% improvement in both CPU and GPU performance are seemingly right on the money, with some fluctuations above and below that figure in various benchmarks and their individual subscores. We can infer that the Snapdragon 845 makes apt use of the architectural improvements provided by the move to A75 and A55 cores, and that the Adreno GPU line once again delivers a respectable year on year improvement. All of this also comes with great power efficiency improvements that, while harder to measure, should result in more-tangible benefits to the end-user. We can also expect performance advantages from the adoption of DynamIQ, one of the more significant developments in ARM-based chipsets recently. Add to that the Snapdragon 845’s shared system cache and the availability of SDKs to make proper use of all SoC blocks, and we can begin seeing how Qualcomm compounded focus on heterogeneous computing will shape the Snapdragon platform moving forward. Tellingly, while the purpose of last week’s press event was primarily to benchmark the Snapdragon 845’s CPU and GPU, most of the tours and talks actually concerned the peripheral components that the company keeps refining with each generation.

Indeed, many of the most exciting developments on Snapdragon lie on the system-on-chip blocks surrounding the CPU and GPU. On the connectivity front, for instance, Qualcomm is improving its modem and working with partners in order to hasten and smoothen the transition to 5G. The company is also doubling down on machine learning, and while its Hexagon 685 DSP falls short of a dedicated processing unit, it still sees three times the performance of the previous generation. The Aqstic audio codec (a low-power audio codec that supports high-resolution standards and integrated DACs), Qualcomm’s power management and fast charging solution, the Spectra ISP, and the new Secure Processing Unit are all value add-ons that impact the user experience in one way or another. Yet, at the same time, it’s been excruciatingly difficult for the company to communicate how all of this extra silicon ultimately works its way into the user experience in concrete, traceable ways. CPUs and GPUs remain the most important components in the minds of most users.

Which leads me to the point I raised in 2016: I noted the widening gap between Apple and Qualcomm, and the ways competitors like Huawei and Samsung were beginning to challenge the company’s performance crown in the Android space. Indeed, that chokehold hasn’t loosened yet — it’s only tightened as the A11 Bionic has leaped ahead of both the Snapdragon 835 and the unreleased 845 in a single revision. As John Poole, creator of Geekbench 4, once said in an interview with XDA: “[A]s much as they’re not competing with Apple, they’re competing with Apple”. This is especially true in the eyes of enthusiasts and those who closely follow mobile technology — it’s becoming increasingly obvious that competitors are catching up, and, in some (or even many) areas, surpassing Qualcomm. With Samsung promising a gigantic twofold increase in single-core performance with its upcoming Exynos chip, for example, and with HiSilicon introducing the first dedicated neural network-specific processing unit last year, much of the press’ attention is being drawn elsewhere.

Sure, Qualcomm will argue that its Hexagon DSP is actually a third-generation AI platform; that their chips are unrivaled in performance per watt, performance per square millimeter, or performance per watt per square millimeter; that they have a larger, broader and more diverse customer base which employs the platform in many different ways; and so on and so forth. These might be solid rebuttals, and I happen to see the validity of some of these talking points. But at the same time, I am of the opinion that the internet at large is still laser-focused on CPU and GPU figures, and the silicon market is only getting fiercer in that realm. That isn’t to say, of course, that Qualcomm’s research and development teams are doing the wrong thing by investing so heavily on all the components that contribute to the user experience, either directly or by allowing OEMs to save costs by adopting standardized implementations such as Quick Charge.

At the end of the day, you probably clicked on this article because you read the word “benchmark” in the title. Looking at our own statistics and the performance of competing sites’ articles on these subjects, I don’t think I’d be wrong to say that you would have been less likely to read an article with a headline about the Aqstic audio codec, the Spectra 280 ISP, the Hexagon 685 DSP, or the Secure Processing Unit. This is one of Qualcomm’s challenges going forward if it is to continue to “only” deliver performance improvements on the order of 30% for the next few years. The widening gap in benchmark scores that the internet claims to care so little or so much about, but in any case can’t seem to stop discussing, will keep siphoning the well-deserved attention that many of the company’s breakthroughs deserve.

If you’re interested in learning more about what the Snapdragon 845 has to offer, check out our past coverage: 


Want more posts like this delivered to your inbox? Enter your email to be subscribed to our newsletter.

Source link