Get ready for smart apps
Best performing Android phones for 2018
This year will finally deliver the benefits of Deep Learning to mobile platforms. We expect significant improvements in privacy, personalization, offline functionality and cost of services across all mobile application segments. Alexa, Cortana and Siri will soon live on our phones, answering questions, translating and being helpful even when we’re traveling abroad or hiking off-the-grid. Video games will become more entertaining, challenging and engaging even when we play against the computer. Video streaming will take less of our bandwidth and mobile data, while the image quality will improve. All of this will be powered with Deep Neural Network technology.
If you are in the market for a new Android phone, we have a few tips to help you choose the best device for this new wave of smart apps.
So, which phone should I get?
At Numericcal, we spend a lot of time evaluating the suitability of various devices for running Deep Neural Networks. To evaluate each phone, we run a battery of Deep Neural Network benchmarks and calculate its Neural Processing Capability (NPC). Intuitively, NPC tells us how much faster a phone runs Deep Neural Networks in comparison to other phones.
If you are optimizing for speed only three brands stand out: Google, Samsung and LG. Surprisingly, the newcomer Google is not only among the top three, but significantly ahead of the second placed Samsung, with a ~20% performance advantage. The following chart shows the top ten phone models according to the raw NPC performance (higher is better). It is interesting to note that the newest model is not always the best performing device.
If you are looking for best value you should consider the following four brands: Samsung, LG, HTC and Motorola. The following chart shows the top ten phone models in terms of NPC per unit price (higher is better).
Despite all the hype, Deep Learning has yet to provide real value for consumers in mobile space. While Deep Neural Network training has been mostly worked out, much work remains to be done to make their deployment and updating on mobile devices seamless and efficient.
In terms of the engineering we see Google, the newcomer to the hardware arena, pulling ahead. It will be interesting to see if incumbents have the necessary technology and expertise to catch up with Google in hardware/software integration.
It is also clear that, as consumers, we will have to pay premium for improved user experience that higher performing hardware can provide. Only four companies made it into top ten phone models according to our performance per unit price metric.
Samsung and LG are featured in both graphs, so it’s fair to say that they currently provide the best overall value. However, even here we must be careful and note that newer models, coming at the premium price point, do not necessarily perform better than the older ones.
I’m a developer and I want to know more
In this performance comparison we used TensorFlow Mobile runtime engine. Of course, there are different ML frameworks and execution engines developers could use for deployment. However, due to the overall ease of use, and the fact that Android ecosystem as well as platform hardware are highly fragmented, TensorFlow Mobile will be the most production ready off-the-shelf DL framework for the next few years.
For the detailed summary of measurement results visit our project workspace. To see how you can leverage these benchmark charts to optimize your app performance in deployment check out our Alpha and let us know what you think.