Snapdragon 835 Hands On and Qualcomm Visit Part 1: Benchmarks, Performance & Power Savings
Last week, we were invited to Qualcomm’s Headquarters in San Diego, California, to get a first look and hands on with the Snapdragon 835 in the flesh.
We were able to put the company’s upcoming chipset through its paces, as well as learn about its product design and philosophy by speaking to project leads and touring the enormous set of Qualcomm offices to learn more about their camera technology, Virtual Reality advancements, and the ways in which they optimized power efficiency. It was an interesting trip that allowed us to get a feel for how the Snapdragon 835 will perform in devices coming this April and beyond, and we got to learn some extra information about what the company is trying to achieve with this new processor; what new features they’re trying to sell to OEMs and consumers alike, and how they intend to market many of these new aspects.
While the core of this trip surrounded benchmarking the Snapdragon 835, Qualcomm highlighted the opinion that far too many mobile enthusiasts miss the forest for the trees by focusing solely on year-on-year performance gains. Admittedly, much of what they wish to convey is hard to measure and quantify, and much harder to meaningfully convey with real-world examples. Nevertheless, we’ll go over some of the things we learned after touching upon what you’ll probably find the most interesting part of this article: benchmarks.
|Specs||Qualcomm Snapdragon 835||Qualcomm Snapdragon 820|
|Chipset||835 (10nm LPE)||821 (14nm LPP)|
|CPU||4x 2.45GHz Kryo 280 (big), 4x 1.9GHz Kryo 280 (LITTLE)||2x 2.15GHz Kryo, 2x 2.19GHz Kryo|
|GPU||Adreno 540 GPU||Adreno 530 GPU at 653MHz|
|Memory||2x 1866MHz 32-bit LPDDR4X||2x 1866MHz 32-bit LPDDR4|
|ISP/Camera||Dual 14-bit Spectra ISP 14-bit 32MP||Dual 14-bit Spectra ISP 25MP|
|Modem||Snapdragon X16 LTE (Cat 16 downlink, Cat 13 uplink)||Snapdragon X12 LTE (Cat 12 downlink, Cat 13 uplink)|
With the official unveiling of the Snapdragon 835 earlier this year, we finally learned about the year-on-year gains the new processor provides over the Snapdragon 820 and 821 through official numbers provided by Qualcomm. Samsung was quick to boast about the performance improvements that their new 10nm FinFET process enables – up to 27% higher performance at the same power usage, or 40% lower power consumption at a similar performance level, while Qualcomm’s numbers were slightly lower at 25% year-on-year boosts for the CPU and GPU. This came as a surprise given that traditionally, Qualcomm itself has cited much higher proportional jumps in performance for their flagship-class releases.
Let’s put it into perspective by comparing it to previous figures – take the Adreno GPU, for example. The Snapdragon 805 was reported to be 40% faster than the Adreno 330 in the 800 and 801, while the Adreno 430 in the Snapdragon 810 further boosted performance by 30%. The Adreno 530 found on the Snapdragon 820 and 821 (with different clockspeeds) offers up to 40% better graphics performance over the previous generation. Now, all of these proportional increases don’t always translate directly into equally-higher benchmark results, and Qualcomm has remained at the top of the graphics game through this steadfast GPU portfolio. But it begs the question, why on Earth did Qualcomm claim a meek 25% figure for this generation? While we’ve learned that the new Adreno revision is just that – a rather slight revision – the CPU itself sees a new architecture, dropping Kryo cores for an ARM-based “semi-custom” core through a licensing agreement, that enables very limited modifications on Qualcomms part (at the event, they were still unwilling to confirm whether the new CPU is based on A72 or A73 cores). What kind of gains can we actually expect from this chipset, then?
We had the opportunity to test the Snapdragon 835 for a short two hours, which was enough time for us to diligently test a variety of benchmarks including Geekbench 4, 3DMark, GFXBench, Basemark OS II, PCMark, and AnTuTu while still allowing the device to reasonably cool down in-between runs, to gather better samples for independent runs. The device the processor was found inside of was an unassuming light weight plastic phablet with a matte body, and top-notch specifications to ensure as few bottlenecks as possible. As per the table below, these include a 1440p display, 6GBs of DDR4 RAM, and fast UFS storage –while Qualcomm wasn’t able to disclose on-site which specific solution they employed here, it was most certainly UFS 2.1 judging from the read and write speeds I was able to achieve using Androbench.
|Device||Qualcomm Test Device|
|Resolution||1400 x 2560|
|Camera||21.4MP / 13MP|
|Storage||64GB UFS (2.1?)|
|Frequency Range||300-2457.6 MHz|
Before we jump to the numbers, I want to point out some caveats you need to know when interpreting these results: the numbers for the Snapdragon 821 and Kirin 960 were obtained through much better-controlled tests with higher sampling, while the limited time only allowed us to gather between three and eight samples per benchmark. The software on the test device was also unstable, and often decided to begin producing terrible results until it was rebooted (we were advised to do so by Qualcomm, as they pointed out this was a bug). We monitored CPU frequency throughout the test and didn’t find anything out of the ordinary which allows us to infer that there was no cheating. Finally, this device featured excellent thermals that peaked at around 33°C (91°F) as measured by our FLIR thermal camera. We wish we could have done more careful testing, and we will definitely take a much deeper look at the 835 once we get our hands on actual devices.
Beginning with CPU performance under Geekbench 4, the test device managed to score an average of 6403 for Multi core and 2040 for single core across 8 independent runs, with the highest score being 6461 for multi-core and 2067 for single core scores. This is a substantial improvement over the Snapdragon 821 that not only is higher than the supposedly-leaked benchmarks we’ve seen circulating on the blogosphere, but also higher than the 25% average would suggest. For reference, our OnePlus 3T (with no benchmark cheating, of course) achieves a mean multi core score of 4344 and 1828 for single core. This means we see over 45% improvements in multi core, but only slightly above 10% for single core. However, there are a few things to consider here: the Snapdragon 835 has an octa-core chip with an asymmetric big.LITTLE setup, while the 821 and Kryo focused on fewer but more-powerful and symmetrical cores.
The multi core year-on-year improvement looks to be substantial, mostly benefitting multi-threaded usage scenarios while still outputting respectable performance for applications which rely on a single core. Surprisingly enough, these scores are also higher than the numbers we obtained for the Kirin 960 in the Huawei Mate 9 (set to “Performance”), scoring a little less than 5% higher in both single and multi core scores. Geekbench 4 itself is one of the better predictors of CPU performance out there, so these results alone are quite revealing, and also provides more clues about the Snapdragon 835’s CPU architecture.
We find a similar story in the GPU department, where 1080p Manhattan Offscreen (ES 3.1) outputs results higher than we expected given Qualcomm’s official numbers. The device offers a 33% year-on-year improvement over the scores we obtained on our Google Pixel XL, and more than 50% the framerate of the G71 in the Kirin 960 (Mate 9). Other tests show similar gains, including 3DMark Slingshot Unlimited 3.1 (which is independent of resolution), where we find gains upwards of 40% over the Google Pixel XL, and over 60% over the Huawei Mate 9. Minimum and maximum frametimes within the test saw healthy variance, with the minimum frametimes on 1080p Manhattan and the exhaustive Car Chase benchmark sitting below the 16.66ms target.
More holistic and comprehensive tests also put the Snapdragon 835 ahead by a respectable margin, though we’d disregard tests like PCMark given their dependency on system optimizations and the huge variance we’ve seen in scores of different devices sharing the same chipset. Benchmarks like Geekbench 4, which get closer to the metal by using the NDK and bypassing the interpreted language overhead, should suffice in giving us an idea of what kind of number-crunching improvements we can expect from these new processors.
I’d also like to remind our readers that these devices were given to us specifically for the purpose of benchmarking, and the hardware had some of the best thermal profiles I’ve seen on a smartphone, so it’s likely that these results will vary with their implementation, and that performance-over-time and other metrics will also be very different from anything we could have encountered here.
While speaking to various Qualcomm representatives and the head of SoC development, I found that an underlying pattern of their talking points revolved around power efficiency. Senior Director Travis Lenier, for example, explained to me that power efficiency was a core goal for the Snapdragon 835 and that while they could have pushed for even higher performance under their configuration, they think they struck a balance that should favor yearly improvements on battery efficiency slightly higher than the yearly performance improvements.
I also suspect that part of Qualcomm’s conservative (in context) yearly improvement numbers come from the fact that many enhancements to the Snapdragon 835’s CPU and GPU, such as better branch prediction or depth rejection for graphics, don’t really shine on most workloads — some smaller additions, like a larger L2 cache for the efficiency cluster, have much more significant improvements to the real-world user experience than one could measure with benchmarks, too. Qualcomm is ultimately confident that the areas in which they focused on, such as virtual reality, provide very respectable battery life savings.
We managed to see such examples during our visit, as we saw a Snapdragon 821 and a Snapdragon 835 being tested for power draw (using tools you yourself can obtain) while running a couple of demos, in real time. The contraption allowed us to see how the current draw varied under the exact same workload for the 821 and 835. Under the virtual reality demo, we saw a current intake difference of 32%, a substantial delta that also comes with a similar boost in performance – many of these improvements don’t come from the GPU either, but rather sensor data processing and specific VR optimizations in the 835. The difference during a very simple camera demo was still a respectable 27%, though the camera was fixed, pointing at a corner with no real activity, so we didn’t get a chance to move the setup.
This sums up part one of our Snapdragon 835 coverage, in the next portion we will focus on all of the aspects that benchmarks cannot measure, yet impact your user experience (and often go beyond performance). As always, keep in mind that none of the numbers above necessarily mean that smartphones running the Snapdragon 835 will offer exceptional performance, though we surely wish they would.
Moreover, with the changes to the architecture of the CPU, some of the features Qualcomm provided in the 821 that enhanced real-world performance, such as the boost-mode (CPU maxing) triggered by opening applications and other user inputs, will not make their way to this new chipset. It’s understandable, given that this is a vastly asymmetric chipset and that specific functionality in particular wouldn’t lend itself to work as well as it did on a quad-core chipsets with homogeneous cores.
But as we said, there are many things that Qualcomm is doing with the Snapdragon 835 that benchmarks simply cannot capture, and two short hours of benchmarks in a small room with a test unit provided by the company certainly don’t tell us all the answers anyway. In a future follow-up article, we will discuss how the overall package has more to offer than raw performance and power saving improvements, and how Qualcomm’s position in the market specifically requires them to offer value past clockspeeds and core counts.