4 ms·
Google breaks AI performance records in MLPerf using TPUv4
- cinntaile 6y agoUnder figure 1 it says "Comparisons are normalized by overall training time regardless of system size, which ranges from 8 to 4096 chips. Taller bars are better." Does this really make sense? The new TPU should have lots of chips and therefore finish training faster, which would make comparing like this kind of pointless? Am I misunderstanding something here?