# 1. Are Neural Networks really running on the phone in AI-Benchmark? Or is it just a simulation?
Yes, exactly the same networks that produced visual results are running on the phone.
# 2. Is AI Benchmark supporting AI Chips available on some recently introduced phones?
Our latest release is supporting phones' AI Accelerators on Android 8.1 and above via NNAPI. On older Android versions, all computations are running on CPU.
# 3. What ML library are you using for training and running Neural Networks?
Currently, Tensorflow is used as a back-end in AI Benchmark. This library was chosen due to its tight integration with Android platform and popularity among research and development ML teams.
# 4. My phone has an AI Chip and Android 8.1+, but the resulting scores are too low. What's missing?
This happens because your phone's manufacturer hasn't provided drivers for its AI accelerator, or made them inaccessible like it happened with Google Pixel phones (i.e. only Google has access to Pixel's AI Chip). Therefore, default Android drivers are used and thus the resulting scores can be even lower than without acceleration. Try waiting for a software update that might bring the appropriate drivers to your device.
# 5. The benchmark finished with an error after the last test. What was wrong?
This is a normal behaviour: in the last test the benchmark is testing how much memory can be allocated to a neural network and thus pushing the phone to its limits. When these limits are reached, the application can be terminated unexpectedly by an operating system. Note that all results (including the last test) are already stored inside the app, so just restart it again to view the final score.
# 6. I am getting an error "TEST FAILED... No enough RAM to run the net." What does this mean?
This means that your phone doesn't have enough free RAM to run a neural network. Try closing all applications running in the background or rebooting the phone. Note that this is a normal behaviour for devices with less than 1GB of RAM: in general, they might not able to run all test from this benchmark. For more information please refer to our ranking table.
# 7. I have trained an awesome network for an awesome task! Can you include it into AI Benchmark?
Send us a link to your paper/technical report and code with pretrained models, and we may consider including it into a new benchmark release. Also, have a look at our current competition.
