Why pi needs so many time?

[samjabrahams]

Why pi needs so many time?

[samjabrahams]

Hi @alephman - I still need to do specific tests, but I believe that the issue has to do with limited RAM, converting 64-bit operations to 32-bit (Raspbian is 32-bit operating system), and/or data transfer speeds. Sorry for the delayed response: lots of work going on.

[samjabrahams]

Just to make a note on this: it turns out that the main culprit in the benchmark is that TensorFlow performs optimizations on-the-fly as you use a model. That is, the first time you load up a graph and run it, it is much slower than subsequent runs. In order to get comparable speed, you need to keep a Session alive in the background instead of loading up a separate Session each time you want to make a run.

I updated the benchmarks at the beginning of December- sorry for the delay in responding to this.