Google announces TPU v4 with MLPerf 0.7 records
In one sentence Posting MLPerf Training 0.7 results, Google reveals TPU v4, a new custom deep-learning accelerator, claiming it built the "world's fastest training supercomputer" with a 4,096-chip pod.
While NVIDIA announces A100, Google answers with a new generation of its own ML-purpose chip: TPU v4. It's not a GPU: it's designed from scratch for the operations AI models need, and individual chips connect into huge clusters ("pods") with very fast links.
Google enters the international MLPerf benchmark and breaks every training-speed record: some models that used to take hours now train in minutes.
For developers on Google Cloud, it means AI hardware competitive with NVIDIA's best, but with a different architecture, will soon be available.
Companies
Tools
TPU v4
Tags
Sources