A new AI accelerator chip is promising dramatic efficiency gains for both training and inference workloads.

Benchmarks suggest meaningful improvements over prior generations.

Cloud providers are expected to integrate the hardware later this year.