CPU Performance
Reference Testing Platform
At the end of the day, Qualcomm allowed me some time with the Snapdragon 810 reference platforms – a pair of devices built around the SD 810 SoC to demonstrate the performance features the processor can offer. One of the units was a tablet while the other was a phone, though quite a bit larger to make room for debugging and monitoring hardware connections. The results that I will be presenting here today are from the tablet device, though I did run benchmarks on both units. As it turned out, the scores from the phone and the tablet hardware were essentially identical so doubling up the scores in our graphs was useless.
Qualcomm has several flagship smartphone design wins for the Snapdragon 810, but I was of course wary of comparing the performance numbers from an 810 tablet to those from currently shipping smartphones that have a much smaller space for cooling and heat dissipation. In theory, a larger device should have more space for a heatsink and could scaling to higher clock speeds for longer periods of time than would possible in a phone. Qualcomm has assured me that the devices were configured exactly as if they were integrated into a phone but you should keep that difference in mind as we walk through the next couple of pages.
I will be comparing the Snapdragon 810 to some of the most recently flagship phones and tablets from the Android market and Apple devices. We have the iPad Air 2 and the iPhone 6 to represent the latest from Apple’s A8 processor line, the international Samsung Note 4 with the Exynos octa-core SoC, a couple of Tegra devices and phones using the Snapdragon 805 and 801 parts.
| Snapdragon 810 Ref | Nexus 6 | OnePlus One | Galaxy Note 4 | SHIELD Tablet | |
|---|---|---|---|---|---|
| SoC | Snapdragon 810 | Snapdragon 805 | Snapdragon 801 | Exynos 5433 | Tegra K1 |
| CPU Cores | Quad-core 2.0 GHz Cortex-A57 Quad-core 1.5 GHz Cortex-A53 |
Quad-core 2.7 GHz Krait 450 | Quad-core 2.5 GHz Krait 400 | Quad-core 1.9 GHz Cortex-A57 Quad-core 1.3 GHz Cortex-A53 |
Quad-core 2.2 GHz Cortex-A15 |
| GPU Cores | Adreno 430 | Adreno 420 | Adreno 330 | Mali-T760 | 192-core Kepler |
| RAM | 4GB LPDDR4-1600 | 3GB LPDDR3-1600 | 3GB LPDDR3-1600 | 3GB LPDDR3-1650 | 2GB LPDDR3-1600 |
| Network | Qulalcomm Cat 9 LTE | Qualcomm MDM9x25 UE Category 4 LTE | Qualcomm MDM9x25 UE Category 4 LTE | Ericsson M7450 Cat.4 LTE | None |
| Connectivity | 802.11a/b/g/n/ac (Wave 2) (2.4/5 GHz) Bluetooth 4.1 USB 3.0 MHL, NFC |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.1 USB 2.0 NFC |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.1 USB 2.0 NFC |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.1 USB 2.0 MHL, NFC |
802.11a/b/g/n (2.4/5 GHz) Bluetooth 4.0 USB 2.0 |
| OS | Android 5.0.2 | Android 5.0.1 | Android 4.4.4 | Android 4.4.4 | Android 5.0.1 |
| Nexus 9 | Dell Venue 8 7000 | iPhone 6 | iPad Air 2 | |
|---|---|---|---|---|
| SoC | Tegra K1 Denver | Atom Z3580 | Apple A8 | Apple A8X |
| CPU Cores | Dual-core 2.3 GHz Denver | Quad-core 2.3 GHz Silvermont | Dual-core 1.4 GHz Cyclone | Triple-core 1.5 GHz Cyclone |
| GPU Cores | 192-core Kepler | PowerVR G6430 | PowerVR GX6450 | PowerVR GX6850 (8-core) |
| RAM | 2GB LPDDR3 | 2GB LPDDR3 | 1GB LPDDR3 | 2GB LPDDR3 |
| Network | None | None | Qualcomm MDM9x25 UE Category 4 LTE | Qualcomm MDM9x25 UE Category 4 LTE |
| Connectivity | 802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.1 USB 2.0 NFC |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.0 USB 2.0 |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.0 USB 2.0 |
802.11a/b/g/n/ac (2.4/5 GHz) Bluetooth 4.0 USB 2.0 |
| OS | Android 5.0.1 | Android 4.4.4 | iOS 8.1.3 | iOS 8.1.3 |
CPU Performance
Geekbench 3 is Primate Labs' cross-platform processor benchmark, with a new scoring system that separates single-core and multi-core performance, and new workloads that simulate real-world scenarios. Geekbench 3 makes it easier than ever to find out if your computer is up to speed. Every test in Geekbench 3 is multi-core aware. This allows Geekbench to show you the true potential of your system. Whether you're running Geekbench on a dual-core phone or a 32-core server, Geekbench is able to measure the performance of all the cores in your system.
Geekbench acts much like a traditional synthetic processor benchmark would, giving us an idea of the peak performance that the CPU offers in both integer and floating point math.
For single threaded integer performance, the Snapdragon 810 falls slightly behind the performance of Samsung’s Exynos 5433 part in the Galaxy Note 4, as well the two Apple processors and NVIDIA’s Tegra K1 Denver. Multi-thread integer results are much more impressive as the SD 810 takes the performance lead, leaving the 8-core Exynos 5433 behind. Compared to the Snapdragon 801 and 805, integer performance scales up by 35% in single and 66% in multi-threaded results.
In floating point testing the Snapdragon 810 jumps over the Samsung SoC in both single threaded and multi-threaded results, but still sits behind the Apple A8X, A8 and Tegra K1 Denver. In multi-threaded floating point performance, the A8X comes out ahead even though it only has 3-cores going against the 8-cores of the Snapdragon 810 and Exynos 5433.
Octane 2.0 is a modern benchmark that measures a JavaScript engine’s performance by running a suite of tests representative of today’s complex and demanding web applications. Octane‘s goal is to measure the performance of JavaScript code found in large, real-world web applications, running on modern mobile and desktop browsers.
The updated Octane 2.0 benchmark includes four new tests to measure new aspects of JavaScript performance, among which: garbage collection / compiler latency and asm.js-style JavaScript performance.
Our testing with Google Octane was done exclusively on the latest version of the Chrome browser on Android, and Safari on iOS.
Qualcomm’s Snapdragon 810 performs well in this first of our three browser-based tests with a score of 8063, falling behind on the Apple A8X and the Tegra K1 Denver processors. Interestingly the Galaxy Note 4, with a very similar A57+A53 core design is 35% slower.
Kraken is a JavaScript performance benchmark created by Mozilla that measures the speed of several different test cases extracted from real-world applications and libraries. The test cases include:
- An implementation of the A* search algorithm
- Audio processing using Corban Brook's DSP.js library
- Image filtering routines, including code from Jacob Seidelin's Pixastic library.
- JSON parsing, including data from Tinderboxpushlog
- Cryptographic routines from the Stanford JavaScript Crypto Library
Our testing with Mozilla Kraken was done exclusively on the latest version of the Chrome browser on Android. and Safari on iOS.
The SD 810 does well in this test too but has a couple more devices best its score of 4618.3ms. Both of the Apple A8 and A8X devices processed the Javascript in less time, as did both of NVIDIA’s Tegra K1 tablets. Notice the large performance jump over both the SD 801 and 805 processors – clearly 810 is going to improve experiences for users.
This is SunSpider, a JavaScript benchmark. This benchmark tests the core JavaScript language only, not the DOM or other browser APIs. It is designed to compare different versions of the same browser, and different browsers to each other.
This test mostly avoids microbenchmarks, and tries to focus on the kinds of actual problems developers solve with JavaScript today, and the problems they may want to tackle in the future as the language gets faster. This includes tests to generate a tagcloud from JSON input, a 3D raytracer, cryptography tests, code decompression, and many more examples. There are a few microbenchmarkish things, but they mostly represent real performance problems that developers have encountered.
This test is balanced between different areas of the language and different types of code. It's not all math, all string processing, or all timing simple loops. In addition to having tests in many categories, the individual tests were balanced to take similar amounts of time on currently shipping versions of popular browsers.
One of the challenges of benchmarking is knowing how much noise you have in your measurements. This benchmark runs each test multiple times and determines an error range (technically, a 95% confidence interval). In addition, in comparison mode it tells you if you have enough data to determine if the difference is statistically significant.
Our testing with SunSpider was done exclusively on the latest version of the Chrome browser on Android, and Safari on iOS.
The Snapdragon 810 scores incredibly well here once again with a time of 542.5ms, beating all of the other Android devices in our suite including the Tegra K1 and Exynos 5433. Only the Apple A8 and A8X are faster but they have the advantage of VERY specific optimizations of Safari for the CPUs.
Vellamo 3.1 is designed to be an accurate, easy-to-use suite of system-level benchmarks for devices based on Android 4.0 forward. In Vellamo we want to enable performance enthusiasts to really understand their system, and how it compares to other systems, and our mission has just begun.
Vellamo began as a mobile web benchmarking tool that today has expanded to include three primary Chapters. The Browser Chapter evaluates mobile web browser performance, the Multicore Chapter measures the synergy of multiple CPU cores, and the Metal Chapter measures the single core CPU subsystem performance of mobile processors.
(Note: the graphs will say 3.0 but 3.1 was run on all setups.) This test is Android-only, so the Apple iPad Air 2 and iPhone 6 results will be ignored.
The first Vellamo test uses Chrome and looks at overall browser performance where the Snapdragon 810 result is 32% faster than SD 805 and 80% faster than SD 801! It still runs slower than the Tegra K1 and Tegra K1 Denver SoCs, but those are significantly higher TDP parts.
Multi-core results attempt to measure how well the different cores can work together to complete computing tasks, and the SD 810 matches the performance of the K1 Denver-based Nexus 9.
Metal is essentially a single-core test and the 810 reference platform pulls a score of 2325, 33% faster than the previous generation Snapdragon.
The cool thing about
The cool thing about UFS–which I just learned–is that it uses the SCSI command stack instead of the SD style of command passing that eMMC uses. So, this should show some small block read/write performance improvements much like UAS did for USB. That could make a big difference. Be sure to test it if a device ever uses such a storage memory, please!
It will definitely be
It will definitely be something we pay attention to going forward.
WOW those SunSpider JS
WOW those SunSpider JS benchmarks, among others for the A8, and A8x really jump out, the A8 having the Cyclone 2 microarchitecture, that we have never really had an indepth review owing to Anand’s departure from anandtech. I’m no great fan of Apple, but the custom wider order superscaluar of the A8/A8X, as well as the Denver microarchitecture from Nvidia, sure perform with their 6, and 6+ IPC respectively, wide execution pipline designs. It’s no wonder the mobile phone makers are going to the 8 core ARM reference design A53/A57 Big/Little designs with their 3 IPC per core narrower superscalar resources. Arm holdings’ newest reference design(A72), if it retains its narrower superscalar design, with only 3 IPC per core, is going to have problems competing on the top tier, but for sure those Custom offerings from Apple, and maybe Nvidia(what’s up with Denver?), as well as AMD’s K12 custom ARM microarchitecture can/will continue to take the high end ARMv8A ISA based market. It’s the high IPC per core custom ARM designs that will continue to give Intel headaches in the mobile arena.
I like your CPU tables, and charts, but could you include the IPC per core figures for all these CPU microarchitectures in the future, and try to get more information on metrics such as reorder buffer size, numbers of floating point/integer pipelines, and other microarchitecture details specific to each makers CPU core.
The custom microarchitectures have much more in the way of execution resources, beyond the standard ARM Holdings’ reference designs.
Those A8/A8X designs(mostly from Apples’ acquisition of P.A. semiconductor, and other IP/company acquisitions) are really top notch, and hopefully AMD’s Jim Keller will be designing such a custom ARMv8a ISA based design(K12) to compete with the Cyclone microarchitecture. The newest ARM holdings reference design A72 if it retains the 3 IPC rate, will have to make up for the deficiency with more cores, and hopefully the cores can be individually power gated, or power gated in groups of 2, for finer power scaling, which would be great for octo-core or higher Tablet/phone SKUs. Apple did a great job of acquiring brainpower, and AMD is on the right track, with Jim Keller, and Lisa Su, and getting AMD a custom ARM ISA based microarchitecture.
Even when the rumors were a
Even when the rumors were a consideration, I am more concerned with the heat generated by displays than the CPUs at this point.
Very interesting article
Very interesting article Ryan. My only concern is that the reference tablet and phone are very large and could be dissipating heat better than, let’s say, a phone with an HTC One chassis?
I know that’s an obvious point and that this is a preview, but it still bugs me. Again, great article 😀
So with ufs will we get more
So with ufs will we get more then 100kb to 1.1mb transfers? I have a old Asus Tf700 and I tell ya transferring anything to or off is super slow. Or transferring from Pc to tablets you can only transfer 1 file at a time so annoying.
Anything and everything “old”ASUS T1700.
The CPU is a NVIDIA® Tegra® 3 Quad-core, 1GB memory, Storage 32GB / 64GB *1 EMMC + 8G life time ASUS Webstorage space, USB 2.0, Wi-Fi a/b/g/n 2.4 Mhz, Storage 32GB or 64GB. Compare those with what this will have…
Anything and everything “old” is going to be slow. Doesn’t matter what it is. USB, Wi-FI, and storage.
Just out of curiosity, I looked up your
I think we will be able to safely consider the Asus T1700 to be a dinosaur one this Qualcomm SoC hits the markets.
So this means the Asus zen 2
So this means the Asus zen 2 is aimed squarely to counter an top of the line 20nm?Intel pretty much set their timetable to counter next gen top of the line with 14 mm?(apple iOS,Samsung)men this will be a interesting 2015