AI for all
here is a preview, it is being filed this week to all app stores, it will be the 1st benchmark that allows total comparaison between Windows on ARM, Windows x86/x64 and iOS and Android, Mac.
Manufacturers are not going to be allowed to optimize the code, as they do on other benchmarks, to bend the reality.
Please feel free to use it
https://apps.apple.com/us/app/mlbenchy/id1482528571
Works on Mac and iPhones/iPads, soon on Windows and Android next year.
Best regards
Francois
There will be a version available as soon as Apple approves it
In case you are wondering how @intel #x86 #MacBookPro stands against #ARM @apple platform on #MachineLearning, use #MLBenchy to do so.
Get it soon in your App Store.
| MacBookPro | iPhone XsMax | Speed up | |
| InceptionV3 | 527 | 132 | 3.99 |
| Nudity | 1355 | 88 | 15.40 |
| Resnet50 | 649 | 102 | 6.36 |
| Car Recognition | 930 | 112 | 8.30 |
| CNNEmotions | 303 | 121 | 2.50 |
| GoogleNetPlace | 847 | 107 | 7.92 |
| GenderNet | 94 | 87 | 1.08 |
| TinyYolo | 558 | 166 | 3.36 |
Those numbers are very trouble some for Intel, of course, Intel will come up with an acceleration (in Progress) for their inference engine, but this shows a piece of the number used by the engineering team of apple to push the decision to switch away from Intel, there is no doubt that they have a firm case to switch.
it is running few neural networks design . The application is pretty large, around 800MB because it includes 7 of the most commonly used inference. We recommend that you only download it over wifi.
Nudity detection model is 224×224, net size is 24MBTinyYolo is 416×416, net size is 64MB
Restnet is 224×224, net size is 100MB
Sentiment is a Dictionary based inference, with a string to Double association , Net size is 275KB.
Car Recognition is 224×224 , net size is 25MB
CNNEmotion recognition is 224×224, net size is around 400MB
GenderNet is 227×227, net size is 45MB
InceptionV3 is 299×299, Net size is 94MB
GoogleNetPlaces is 224×224, and net size is 25MB
1st iteration of the benchmark is loading the neuronal network into memory, while running the inference once. It shows you the warm up time of the system.
2nd iteration is having too much variation to be meaning full.
3rd iteration shows you a pretty stable number.
Please contact us if you need interpretation assistance.
Notice that a PC version will be coming, to compare the Apple capabilities against PC AI capabilities.