Google Makes Cloud TPU Accelerators Available to Public Beta Users

Google opens up its custom Tensor Processing Units on the Google Cloud Platform, allowing users to run machine learning workloads without specialized hardware expertise. The beta release provides access to powerful ASICs designed specifically to accelerate TensorFlow applications.

Google launches its custom Tensor Processing Units (TPUs) into public beta on the Google Cloud Platform, giving developers direct access to hardware built specifically for machine learning. These second-generation Cloud TPUs deliver up to 180 teraflops of floating point performance and include 64 GB of High Bandwidth Memory to handle demanding artificial intelligence tasks.

Users connect to these accelerators using high-level TensorFlow APIs, eliminating the need for the deep specialized expertise traditionally required to program custom ASICs. Google open-sources several reference models, such as ResNet-50 for image classification and Transformer for machine translation, to help developers start building and training their applications immediately.

The company designs these Cloud TPUs to handle massive consumer-facing workloads ranging from search and voice recognition to self-driving vehicle training. Multiple TPU boards link together via a dedicated high-speed network to form TPU pods, creating multi-petaflop supercomputers that offer an alternative to traditional CPUs and Nvidia GPUs in the cloud.

Read More at the original source →