google-deepmind/alphatensor

Fault Tolerance aspect

Muhtasham opened this issue · 2 comments

Muhtasham commented

Dear Authors,
First of all thanks for the great work and congrats on breakthrough!

In the paper the Fault Tolerance aspect of multiplication algorithms are left out, could you please shed some light on it. I believe this has serious implications.

Sciroccogti commented

Isn't this an exact multiplication algorithms?

  • 👍4
matejbalog commented

Thank you!

Yes, the algorithms in this work are exact.

For testing numerical precision when multiplying matrices of floats, you can see benchmarking/test_correctness.py and the metrics referenced in Supplement B.3 of the paper