XiaoMi / mobile-ai-bench

Benchmarking Neural Network Inference on Mobile Devices
Apache License 2.0
353 stars 57 forks source link

Kirin970 GPU much slower than Snapdragon 845 #23

Closed achigeor closed 5 years ago

achigeor commented 5 years ago

Are there any insights as to why kirin 970 GPU is that much slower than Snapdragon 845? The compute capabilities should be around the same for these two SoCs right?

For example mobilenetsv2 benchmark runs at ~10ms for sdm845 and 47.597ms for kirin970 (optimized). I also see big differences in my custom model. On a sdm845 I get around 75ms, while on kirin980 I get around 250ms.

Is it because snapdragon is actually that much faster, or because of how Mace ops / kernels are implemented? If it's the latter, what would be a good place to start working on possible optimizations? Would you expect improvements if ARM Compute is used for ARM SoCs?

llhe commented 5 years ago

@achigeor See the reply here: https://github.com/XiaoMi/mace/issues/1#issuecomment-462580008

Welcome to post the results for ARM Compute Library to make the comparison, and adding an executor will be highly appreciated.