MLPerf Training 0.7: Google records on TPU v3 and a first look at TPU v4
In one sentence In the MLPerf Training v0.7 results, Google sets records in 6 of 8 benchmarks with a 4,096-chip TPU v3 supercomputer and shows TPU v4 for the first time, on average 2.7x faster than TPU v3 at the same chip count.
MLPerf is a benchmark contest where hardware makers measure how long it takes to train a set of reference models. In the July 2020 round, Google shows up with a 4,096-chip TPU v3 supercomputer, built from its own accelerators designed specifically for machine learning.
The result: records in 6 of the 8 benchmarks, with several models trained from scratch in under 30 seconds.
In the same round Google also shows the next generation, TPU v4, not yet available to customers: at the same chip count it is on average 2.7 times faster than TPU v3.
Companies
Tools
TPU v3, TPU v4
Tags
Sources