Google Opens Custom TPU Machine Learning Chips to Public Beta

Google makes its custom Tensor Processing Units available in beta, offering faster and more efficient machine learning performance for TensorFlow users. The service aims to give Google Cloud a unique edge over competitors like AWS and Azure.

Google releases its custom Tensor Processing Units (TPUs) to developers in beta, providing specialized hardware designed to accelerate machine learning workloads in TensorFlow. These custom chips deliver significantly faster performance than standard GPUs, reaching a peak of 180 teraflops per board, while consuming less power to keep operational costs down.

Each Cloud TPU board features four custom ASICs paired with 64 GB of high-bandwidth memory. Developers who already build in TensorFlow can utilize these chips without rewriting their code, though they currently have to request a quota and describe their use case before gaining access. The service costs $6.50 per TPU per hour, which compares to $1.46 per hour for a less powerful standard GPU.

This launch gives Google Cloud a major differentiator against competitors like AWS and Azure in an era where basic cloud services and container technology make it easy to switch providers. By pairing its popular TensorFlow framework with exclusive TPU hardware, Google creates a powerful ecosystem that rivals cannot easily match in the short term.

Read More at the original source →