More options for customized chips for learning and reasoning
2027 expected to expand in phases
사진 확대
Google’s semiconductor tensor processing unit (TPU) dedicated to artificial intelligence (AI) is provided through Oracle Cloud Infrastructure (OCI). This is the first time Google TPU will be offered in a hyperscale cloud rather than Google. Google TPU will be available from 2027 on OCI in the future.
Oracle and Google announced a partnership plan on the 28th (local time).
The background of the two companies’ partnership is the change in demand for AI computation. As the model size grows and the demand for inference in real service environments increases, the overall performance and cost-effectiveness of learning and inference are becoming important.
TPU is a customized AI accelerator designed by Google and is integrated with open software. It supports the entire process of AI development and operation, from state-of-the-art model learning, large-scale inference, and multi-level inference of agentic AI.
Google explained that it designed the TPU to increase AI computational throughput per watt of power through a structure that allows fast access to model data and high-speed communication between chips.
TPU supports the learning and post-learning of large-scale models such as dense (dense) models and expert mixed (MoE) models. MoE is a structure that reduces the amount of computation by selecting and operating only a part of several ‘expert’ submodels depending on the question. It is also used to infer inference AI and agentic AI, and long context services that process long documents or conversations at once. OCI customers can choose TPU to meet their workload, performance needs, and cost goals.
Mahesh Thiagarayan, president of Oracle OCI, said, “Each AI workload has different computing requirements,” adding, “The introduction of Google TPU in OCI will allow customers to use multiple key AI computing architectures in one cloud.” “By securing this range of choices, we can flexibly select the right infrastructure for the performance, throughput, and affordability of each AI workload,” he explained.
“The next generation of AI will require different types of computing power,” said Amin Bachdat, senior technical director and fellow at Google Cloud AI Infrastructure. “TPU provides customers with a powerful and dedicated designed option to handle demanding training and inference workloads.” “The introduction of Google TPU in OCI will not only allow more organizations to utilize the technology, but will also provide greater flexibility in building and running the most sophisticated AI applications,” he added.