Machine Learning Inference Performance

AIMark 3

AIMark makes use of various vendor SDKs to implement the benchmarks. This means that the end-results really aren’t a proper apples-to-apples comparison, however it represents an approach that actually will be used by some vendors in their in-house applications or even some rare third-party app.

鲁大师 / Master Lu - AIMark 3 - InceptionV3 鲁大师 / Master Lu - AIMark 3 - ResNet34 鲁大师 / Master Lu - AIMark 3 - MobileNet-SSD 鲁大师 / Master Lu - AIMark 3 - DeepLabV3

In AIMark 3, the benchmark uses each vendor’s proprietary SDK in order to accelerate the NN workloads most optimally. For Qualcomm’s devices, this means that seemingly the benchmark is also able to take advantage of the new Tensor cores. Here, the performance improvements of the new Snapdragon 865 chip is outstanding, posting in 2-3x performance compared to its predecessor.

AIBenchmark 3

AIBenchmark takes a different approach to benchmarking. Here the test uses the hardware agnostic NNAPI in order to accelerate inferencing, meaning it doesn’t use any proprietary aspects of a given hardware except for the drivers that actually enable the abstraction between software and hardware. This approach is more apples-to-apples, but also means that we can’t do cross-platform comparisons, like testing iPhones.

We’re publishing one-shot inference times. The difference here to sustained performance inference times is that these figures have more timing overhead on the part of the software stack from initialising the test to actually executing the computation.

AIBenchmark 3 - NNAPI CPU

We’re segregating the AIBenchmark scores by execution block, starting off with the regular CPU workloads that simply use TensorFlow libraries and do not attempt to run on specialized hardware blocks.

AIBenchmark 3 - 1 - The Life - CPU/FP AIBenchmark 3 - 2 - Zoo - CPU/FP AIBenchmark 3 - 3 - Pioneers - CPU/INT AIBenchmark 3 - 4 - Let's Play - CPU/FP AIBenchmark 3 - 7 - Ms. Universe - CPU/FP AIBenchmark 3 - 7 - Ms. Universe - CPU/INT AIBenchmark 3 - 8 - Blur iT! - CPU/FP

Starting off with the CPU accelerated benchmarks, we’re seeing some large improvements of the Snapdragon 865. It’s particularly the FP workloads that are seeing some big performance increases, and it seems these improvements are likely linked to the microarchitectural improvements of the A77.

AIBenchmark 3 - NNAPI INT8

AIBenchmark 3 - 1 - The Life - INT8 AIBenchmark 3 - 2 - Zoo - Int8 AIBenchmark 3 - 3 - Pioneers - INT8 AIBenchmark 3 - 5 - Masterpiece - INT8 AIBenchmark 3 - 6 - Cartoons - INT8

INT8 workload acceleration in AI Benchmark happens on the HVX cores of the DSP rather than the Tensor cores, for which the benchmark currently doesn’t have support for. The performance increases here are relatively in line with what we expect in terms of iterative clock frequency increases of the IP block.

AIBenchmark 3 - NNAPI FP16

AIBenchmark 3 - 1 - The Life - FP16 AIBenchmark 3 - 2 - Zoo - FP16 AIBenchmark 3 - 3 - Pioneers - FP16 AIBenchmark 3 - 5 - Masterpiece - FP16 AIBenchmark 3 - 6 - Cartoons - FP16 AIBenchmark 3 - 9 - Berlin Driving - FP16 AIBenchmark 3 - 10 - WESPE-dn - FP16

FP16 acceleration on the Snapdragon 865 through NNAPI is likely facilitated through the GPU, and we’re seeing iterative improvements in the scores. Huawei’s Mate 30 Pro is in the lead in the vast majority of the tests as it’s able to make use of its NPU which support FP16 acceleration, and its performance here is quite significantly ahead of the Qualcomm chipsets.

AIBenchmark 3 - NNAPI FP32

AIBenchmark 3 - 10 - WESPE-dn - FP32

Finally, the FP32 test should be accelerated by the GPU. Oddly enough here the QRD865 doesn’t fare as well as some of the best S855 devices. It’s to be noted that the results here today were based on an early software stack for the S865 – it’s possible and even very likely that things will improve over the coming months, and the results will be different on commercial devices.

Overall, there’s again a conundrum for us in regards to AI benchmarks today, the tests need to be continuously developed in order to properly support the hardware. The test currently doesn’t make use of the Tensor cores of the Snapdragon 865, so it’s not able to showcase one of the biggest areas of improvement for the chipset. In that sense, benchmarks don’t really mean very much, and the true power of the chipset will only be exhibited by first-party applications such as the camera apps, of the upcoming Snapdragon 865 devices.

System Performance GPU Performance & Power
Comments Locked


View All Comments

  • iphonebestgamephone - Tuesday, December 31, 2019 - link

    "I am into android from the start + symbian before than and also senior member with dev/helping known devs with project @ xda. So thank you, I know enough about android." Haha... I should have known you would come up with something like that.

    "Btw, used android for 10 years (only high end phones) till I switched to the pro max + I have highly technical background as education, hobby and work - especially in the field of electronics and computers" this one too lol.

    And then you somehow decide civ6 and deadcells dont run cus android too weak. No. Its just the devs dont bother with it. They could have restricted it to atleast sd820 devices like what grid autosport devs are doing.

    "Emulation is cool, did a lot on android with it. Including fun stuff like running diablo 2 LOD latest patch on my note 9, believe me - it's playable with the spen when on the go, in home one mouse and the TV = you are good to go. Still, ported or developed games for mobile just works better and you have such a vast library nowdays with high quality games that you really don't need to revisit old classics on your phone. Actually on ios the situation is a lot better, you got a lot more paid apps there vs android." Im yet to find some good stuff like god of war, nfs, burnout, wipeout, xenoblade, pokemon, zelda or mario, on android, or any other thousands of games. You could say you can stream them, but same goes for pc games too. Emulators and a switch style gamepad is great on the go. I see apple has done a great job with metal, vulkan is worse than opengl on android 10 sd855. Looking forward to the updatable drivers on the 865.

    "I can play fortnite maxed at 60fps and no fps drops or whatever even after 2 hours of play without major heating and you are talking about PUBG maxed. :)"
    Thats awesome, sd855 heats up a lot on pubg maxed. I guess there is no pubg gfxtool for ios.
  • cha0z_ - Thursday, January 2, 2020 - link

    There are emulators for ios and you don't need jailbreak to install/play games. They are not on the app store tho, they are on custom stores - still, it's not any different than installing APK from outside playstore. The emulators library is quite big, including ppsspp. As I said tho - android is better for emulators imho + I didn't say android is weak as OS. Weak are the SOCs on android phones compared to the A series of apple. I would totally love to see android phone with apple SOC/similar performance to it and longer full support than two years.

    As for the gfxtool, I hope you understand that when you have literally just a few phones to optimise for - you really do a great job with it, or with other words - the ios pubg variant is greatly optimised for every iphone that supports it to extract the best experience with the best possible gfx for the hardware. Ofc you can argue than personal preferences can apply and tweaking can be done, but it's not that necessary.

    I respect your opinion and share few viewpoints, just from personal experience - gaming on ios is generally better. Hard to explain, games run smoother and better. If you love emulators tho - android is obviously a better choice + snapdragon SOC.
  • iphonebestgamephone - Tuesday, January 7, 2020 - link

    "Weak are the SOCs on android phones compared to the A series of apple." Yeah everyone knows. They still are strong enough for the games you mentioned though, atleast the last 2 years of flagships. And last years 730/730g are also good enough. I guess the devs want even those with 100$ phones play their games. I doubt those people would even bother buying the game once it hits the store.

    Gamebench did a test and the huawei mate 30 pro actually performed better than the iphone 11 pro amd note 10 in games.

    The iphone probably had better visual settings/higher resolution as default probably.
  • Ahmedrr1 - Sunday, December 22, 2019 - link

  • AceMcLoud - Sunday, December 22, 2019 - link

    Ouch, that doesn't look very promising.
  • ballsystemlord - Friday, February 7, 2020 - link

    Spelling error:

    "The test here is mostly sensible to the performance scaling of the A55 cores. The QRD865 in the default more is more conservative than some existing S855 devices,"
    "mode" not "more":
    "The test here is mostly sensible to the performance scaling of the A55 cores. The QRD865 in the default mode is more conservative than some existing S855 devices,"
  • Hrel - Tuesday, April 28, 2020 - link

    Man these Watt listings make no sense at all.

    5.12 Watts is shows as lower than 4.24 Watts then 2.73W is somehow HIGHER than that?! WTF is going on?

    Then 3.33W is higher than 2.73, which makes sense, but then 3.05W is lower than 2.56W?! What are these charts?
  • Hrel - Tuesday, April 28, 2020 - link

    Oh, the bar is for the Joules, the Watts aren't visually represented. Runtime being a critical variable, I gotcha now. Lol, I was so confused :)

Log in

Don't have an account? Sign up now