Google Unveils Third-Generation TPU to Fuel Cloud AI Ambitions
Google introduces its third-generation Tensor Processing Unit, delivering eight times the performance of the previous model to bolster Google Cloud. The release highlights a booming industry race among tech giants and startups to build custom AI chips.
Google reveals its third-generation Tensor Processing Unit (TPU) at Google I/O 2018, bringing eight times more power than last year's model to reach up to 100 petaflops in performance. This new chip is specifically optimized for Google's TensorFlow machine learning framework as the company pushes to make Google Cloud a dominant platform. To manage the immense heat generated by this powerful hardware, Google implements liquid cooling in its data centers for the first time.
The release highlights a fierce industry race where major tech companies design custom silicon tailored to their specific software and operational needs. For example, Facebook develops hardware optimized for its Caffe2 framework to process massive user information graphs, while Amazon builds its own chips to strengthen its AWS ecosystem. Providing specialized machine learning tools is now a basic requirement for any company hoping to compete in the cloud infrastructure market.
Beyond the tech giants, a well-funded startup ecosystem emerges to challenge Nvidia's dominance in machine learning hardware. Startups like Cerebras Systems, SambaNova Systems, and Mythic each raise significant capital to create specialized chips that aim to outperform established products on price or efficiency. As Google already looks ahead to its next generation of hardware, the overall industry rapidly accelerates its shift toward specialized AI processing.