Google Cloud Launches Cloud TPU Accelerators in Beta for Faster ML Training

Google introduces Cloud TPUs in beta on Google Cloud Platform to help machine learning experts train and run their models significantly faster. These custom hardware accelerators feature four ASICs that deliver up to 180 teraflops of floating-point performance and 64 GB of high-bandwidth memory on a single board. Users can access these processors interactively through a Google Compute Engine virtual machine that they fully control and customize.

These powerful boards can operate independently or connect via an ultra-fast dedicated network to create multi-petaflop ML supercomputers known as TPU pods. This setup allows teams to train several variants of a business-critical model overnight instead of waiting days or weeks, enabling them to deploy the most accurate version the very next day. As an example, users can train the ResNet-50 model on the ImageNet benchmark in under a day for less than $200.

Google simplifies the programming process by allowing developers to use high-level TensorFlow APIs rather than requiring specialized expertise in custom ASICs. The company also provides open-source reference implementations for popular models like ResNet-50 for image classification, Transformer for machine translation, and RetinaNet for object detection. Google continuously tests these implementations to ensure they achieve both high performance and accurate results on standard datasets.

Read More at the original source →