GPU Performance & Power

On the GPU side of things, Qualcomm has long been leading the benchmark charts with the help of their in-house Adreno GPU architecture. With the Galaxy S10, we again see a new round of Adreno vs Mali in the Snapdragon and Exynos variants of the phone.

The Adreno 640 in the Snapdragon 855 has relatively conservative performance targets this generation. Here Qualcomm promises 20% better performance even though the GPU itself has a reported 50% more execution units. What has happened is that Qualcomm has dropped the clock frequency from 710MHz down to 585MHz, account for where most of that theoretical GPU performance is missing. The rationale here is to be able to run wider and slower, and thus more efficiently.

On the Exynos side of things, the new chip adopts a new Mali G76MP12 GPU clocked in at up to 702MHz. We’re already seen the GPU inside of the Kirin 980, however for whatever reason Samsung S.LSI has always been able to achieve better results than HiSilicon for several generations in a row, so it’ll be interesting to see how these two chipsets differ.

Starting off with the 3Dmark Sling Shot Extreme Unlimited test suite, the Physics workload is mostly a CPU bound test within a GPU thermally constrained scenario.

3DMark Sling Shot 3.1 Extreme Unlimited - Physics

The Exynos 9820 surprisingly takes the performance lead between both Galaxy S10 units. The result here is a very big change compared to previous generation Exynos SoCs. I hadn’t had the time to investigate if this is actually caused by improvements of the new M4 core or if the workload is being scheduled on the A75 cores. Both peak performance and sustained performance here are very good and are only beaten by Kirin 980 devices.

The Snapdragon 855 Galaxy S10 also posts excellent peak perf results, however the CPU seem to throttle quite a bit more, falling in line with what last year’s Snapdragon 845 devices were scoring.

3DMark Sling Shot 3.1 Extreme Unlimited - Graphics 

In the Graphics score of the workload, we come back to the familiar dominance of Qualcomm GPUs. What is interesting to see here is that both Galaxy S10 units sport worse sustained performance than the Note9 with last year’s chipsets. Most likely this is due to different thermal limits on these two Samsung devices.

GFXBench Aztec Ruins - High - Vulkan/Metal - Off-screen GFXBench Aztec Ruins - Normal - Vulkan/Metal - Off-screen

In the new GFXBench Aztec Ruins tests, the Exynos unit takes the lead in terms of sustained performance in the High variant test, only beaten by Apple’s newest iPhones. The phone doesn’t seem to reproduce the same lead in the Normal variant and subsequently slightly trails the Snapdragon 855 version. In sustained performance, the Exynos S10 beats last year’s predecessors, however the Qualcomm chip merely matches some of the better Snapdragon 845 devices from last year.

GFXBench Manhattan 3.1 Off-screen

In Manhattan 3.1 Off-screen, we see both S10’s neck-in-neck in peak performance, and sustained performance also doesn’t seem all that different. The Exynos variant again shows big leaps over last year’s G72MP18 GPU, and the Qualcomm variant again is only able to match or actually lose out to some of the more thermally aggressive Snapdragon 845 units from last year.

GFXBench Manhattan 3.1 Offscreen Power Efficiency
(System Active Power)
  Mfc. Process FPS Avg. Power
(W)
Perf/W
Efficiency
iPhone XS (A12) Warm 7FF 76.51 3.79 20.18 fps/W
iPhone XS (A12) Cold / Peak 7FF 103.83 5.98 17.36 fps/W
Galaxy 10+ (Snapdragon 855) 7FF 70.67 4.88 14.46 fps/W
Galaxy 10+ (Exynos 9820) 8LPP 68.87 5.10 13.48 fps/W
Galaxy S9+ (Snapdragon 845) 10LPP 61.16 5.01 11.99 fps/W
Huawei Mate 20 Pro (Kirin 980) 7FF 54.54 4.57 11.93 fps/W
Galaxy S9 (Exynos 9810) 10LPP 46.04 4.08 11.28 fps/W
Galaxy S8 (Snapdragon 835) 10LPE 38.90 3.79 10.26 fps/W
LeEco Le Pro3 (Snapdragon 821) 14LPP 33.04 4.18 7.90 fps/W
Galaxy S7 (Snapdragon 820) 14LPP 30.98 3.98 7.78 fps/W
Huawei Mate 10 (Kirin 970) 10FF 37.66 6.33 5.94 fps/W
Galaxy S8 (Exynos 8895) 10LPE 42.49 7.35 5.78 fps/W
Galaxy S7 (Exynos 8890) 14LPP 29.41 5.95 4.94 fps/W
Meizu PRO 5 (Exynos 7420) 14LPE 14.45 3.47 4.16 fps/W
Nexus 6P (Snapdragon 810 v2.1) 20Soc 21.94 5.44 4.03 fps/W
Huawei Mate 8 (Kirin 950) 16FF+ 10.37 2.75 3.77 fps/W
Huawei Mate 9 (Kirin 960) 16FFC 32.49 8.63 3.77 fps/W
Huawei P9 (Kirin 955) 16FF+ 10.59 2.98 3.55 fps/W

Looking at the power consumption and efficiency tables in Manhattan 3.1, we see both devices showcase quite similar characteristics. Performance is very close in both chipsets, with also very similar power consumption within 220mW of each other. The efficiency also is quite close to each other. Interestingly both Qualcomm and Samsung weren’t able to close the gap to Apple’s latest iPhones and the A12 which still has a considerable performance and power efficiency lead.

For the Exynos chipset, it’s also unfortunate to see that absolute power has gone up by 1W, meaning the device will heat up faster, even though performance and efficiency is better.

GFXBench T-Rex 2.7 Off-screen

In T-Rex we again see both chipsets perform very similarly with similar sustained performance figures.

GFXBench T-Rex Offscreen Power Efficiency
(System Active Power)
  Mfc. Process FPS Avg. Power
(W)
Perf/W
Efficiency
iPhone XS (A12) Warm 7FF 197.80 3.95 50.07 fps/W
iPhone XS (A12) Cold / Peak 7FF 271.86 6.10 44.56 fps/W
Galaxy 10+ (Snapdragon 855) 7FF 167.16 4.10 40.70 fps/W
Galaxy S9+ (Snapdragon 845) 10LPP 150.40 4.42 34.00 fps/W
Galaxy 10+ (Exynos 9820) 8LPP 166.00 4.96 33.40fps/W
Galaxy S9 (Exynos 9810) 10LPP 141.91 4.34 32.67 fps/W
Galaxy S8 (Snapdragon 835) 10LPE 108.20 3.45 31.31 fps/W
Huawei Mate 20 Pro (Kirin 980) 7FF 135.75 4.64 29.25 fps/W
LeEco Le Pro3 (Snapdragon 821) 14LPP 94.97 3.91 24.26 fps/W
Galaxy S7 (Snapdragon 820) 14LPP 90.59 4.18 21.67 fps/W
Galaxy S8 (Exynos 8895) 10LPE 121.00 5.86 20.65 fps/W
Galaxy S7 (Exynos 8890) 14LPP 87.00 4.70 18.51 fps/W
Huawei Mate 10 (Kirin 970) 10FF 127.25 7.93 16.04 fps/W
Meizu PRO 5 (Exynos 7420) 14LPE 55.67 3.83 14.54 fps/W
Nexus 6P (Snapdragon 810 v2.1) 20Soc 58.97 4.70 12.54 fps/W
Huawei Mate 8 (Kirin 950) 16FF+ 41.69 3.58 11.64 fps/W
Huawei P9 (Kirin 955) 16FF+ 40.42 3.68 10.98 fps/W
Huawei Mate 9 (Kirin 960) 16FFC 99.16 9.51 10.42 fps/W

In the power end efficiency tables we however see a big difference between the two devices. Here Qualcomm is able to clearly achieve lower power and higher efficiency than the Exynos.

One thing that I note on both Galaxy S10 units is that I again saw some very odd thermal behaviour on the part of the Qualcomm unit. Just like we measured on the Note9 a few months ago, the Qualcomm Galaxy S10+ reached much higher initial temperatures than the Exynos S10+. I measured peak skin temperatures on the front screen near the SoC nearing 49°C on the S855 unit while the E9820 peaked around 43°C. Again, much like last year, this seems to be a time-bound boost mechanism as after a certain period of around 20 minutes the Snapdragon unit throttles down to a sustained 42-43°C. What this means is that the Snapdragon unit has higher (longer) peak performance figures at a cost of a hotter device, before both devices equalise at a sustained ~42°C.

Overall, the Snapdragon unit this year does still have a performance and efficiency lead, however the gap has been narrowed compared to what we’ve seen in the past years. The new Mali G76 looks to have made solid improvements, and ALU heavy workloads in particular have seen very large leaps compared to the Exynos 9810.

The Adreno 640 this generation just seems quite conservative – Apple has taken Qualcomm’s performance crown in mobile and most importantly also the efficiency crown. Both the Snapdragon 855 and A12 are both manufactured on the same process node so it’s a valid Dragon-to-Apples comparison, and here Qualcomm is beaten by such a significant margin of which in the past we’ve only been used to seeing Qualcomm beat Arm with. For the next generation, we thus hope both Qualcomm and Arm will be able to show more significant jumps in both performance and efficiency. Samsung’s own GPU is also a wildcard, however I’m not expecting to see this productised in next year’s Exynos.

System Performance Display Measurement
Comments Locked

229 Comments

View All Comments

  • Irish910 - Saturday, March 30, 2019 - link

    Where is that website link that shows how many MORE times android tracked someone over iOS over the course of a day.....
  • id4andrei - Sunday, March 31, 2019 - link

    Google is pretty transparent about what it gathers and how it is using it. You can download at anytime anything relating to your metdata. You can wipe history of that data. You can disable tracking, personalized advertising and more. These controls are available to you in Android and your google account. This is one thing. Saying that Android has ads and has inherent security issues is another and it's plain bullshit. Saying that the Google store is the wild west is also bullshit.
  • name99 - Friday, March 29, 2019 - link

    "Exynos 9820 is the first tri-CPU cluster/group SoC which actually consists of three different CPU microarchitectures"

    It's not exactly comparable, but the A12/A12X has (at least) three different ARMv8 cores on it, the big cores (Vortex), the small cores (Tempest) and the tiny controller cores (but still ARMv8) Chinook. There are doubtless some number of M0s and suchlike ARMv7 cores also scattered around, but it's interesting that there are three different Apple-designed cores.

    It's also interesting, in terms of area, to scan
    https://en.wikipedia.org/wiki/Apple-designed_proce...

    Notable comparisons, for example, are A9X vs A10 (nominally both 16mm FF, but A10 uses the resources more efficiently) and A10 vs A10X.
  • Andrei Frumusanu - Friday, March 29, 2019 - link

    Oh come on you know better than this. The Chinooks are not part of the CPU cluster and aren't userspace program visible.

    I'm also not counting the Cortex A5's in the Exynos' audio and ISPs or the multitude of Cortex M3s it has.
  • name99 - Friday, March 29, 2019 - link

    Don't want to argue about it; I just thought this was an interesting point :-)

    I'd be just as interested if we learned that QC (or ARM proper) were using ARMv8 devices (ie "interesting" cores, not tiny cores) to handle any of their "controller" type functionality, eg controlling the NPU or GPU.
  • tuxRoller - Friday, March 29, 2019 - link

    I've not finished the article so perhaps you address these issues elsewhere.

    "I wish Samsung at least would mimic the haptics with the fingerprint sensor."

    Coupling haptics and an ultrasonic sensor that also looks beyond surfaces seem like it would be more difficult than just measuring capacitance.

    Also, since the ultrasonic sensor works when the screen is off one should expect the apparent interaction time to go up. Did you happen to time it when the screen was on? The last scenario would be timing its unlock cycle when the phone is in use (any of the password managers should be fine).
  • Andrei Frumusanu - Friday, March 29, 2019 - link

    > Did you happen to time it when the screen was on?

    I didn't do high-speed camera testing of it, but it does look every so slightly faster to respond.
  • tuxRoller - Saturday, March 30, 2019 - link

    That's not too bad then. Most of the reviews I've seen haven't mentioned the new sensor being particularly slow, so, your experience stood out to me.
  • name99 - Friday, March 29, 2019 - link

    I don't know if the iOS 12.2 update had a change to scheduler or JS that has an important effect on web scheduling, but I got 124 for Jetstream 2 on my iPhone XS which is, of course, substantially better than the 98 that Andrei sees.

    FWIW I got a very similar number on my iPad Pro A12X, and on my iMac Pro (Xeon W turbos to 4.2 GHz) I got 142, which is remarkably close to the A12/A12X number...
  • tipoo - Sunday, March 31, 2019 - link

    How nuts is it that for largely ST bound tasks like Javascript, the A12 hangs right in there with the Xeon W, which turbos to 4.2GHz.

    Scale up the core count and memory bandwidth and I don't see why anyone would assume that wouldn't be a very competent chip even for higher end systems, if the software for ARM support was there.

Log in

Don't have an account? Sign up now