In 2023, Google announced the creation of Cloud TPU v5p, its most powerful TPU yet, and an AI Hypercomputer from Google Cloud. Amin Vahdat, Google’s Engineering Fellow and Vice President for the Machine Leaning, Systems, and Cloud AI team, noted the increased demands for training, tuning, and inference due to the exponential growth in generative AI models.
The Cloud TPU v5p serves as an AI accelerator, designed to handle large models with long training periods, primarily consisting of matrix computations and no custom operations within its main training loop. It boasts 8,960 chips in each TPU v5p pod using Google’s highest-bandwidth inter-chip interconnect, and surpasses previous iterations with its enhanced FLOP performance and scalability.
Google also introduced the AI Hypercomputer, integrating open software, performance-optimized hardware, machine learning frameworks, and flexible consumption models to improve productivity and efficiency in AI computing. The performance-optimized hardware leverages Google’s Jupiter data center network technology.
In a departure from its usual practices, Google is offering open software with extensive support for machine learning frameworks like JAX, PyTorch, and TensorFlow alongside the AI Hypercomputer. The AI Hypercomputer also introduces two new models, Flex Start Mode and Calendar Mode.
Alongside these developments, Google unveiled Gemini, its “largest and most capable” AI model, and announced its rollout to Bard and the Pixel 8 Pro in three sizes: Gemini Pro, Gemini Ultra, and Gemini Nano.
