Oracle plans to offer Google TPUs on its cloud platform in 2027

Oracle and Google announced on 27 October that Google’s eighth-generation Tensor Processing Units will be available through Oracle Cloud Infrastructure in 2027, making OCI the first hyperscale cloud outside Google to offer the chips, Oracle said. The companies announced the plan at Oracle AI World in Las Vegas. It would give OCI customers a choice between Google’s custom AI accelerators and the GPUs already offered on the platform.

Oracle adds TPUs alongside Nvidia hardware

The planned addition broadens Oracle’s chip range. Oracle has signed deals worth tens of billions of dollars for Nvidia hardware in its data centres, The Register reported. Oracle also announced the general availability of Nvidia Vera Rubin NVL72 systems and Vera CPUs at its Las Vegas event, according to the publication.

Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said the Google arrangement would let customers select among AI compute architectures within one cloud. Oracle says customers will be able to choose TPUs according to the performance and cost requirements of individual workloads, including model training and production inference.

Oracle’s stated role is to provide the power, cooling and data-centre networking needed to deploy the TPUs on OCI. Google Cloud and Oracle said the deployment would begin in calendar year 2027. The Register reported that Oracle gave no quantity for the TPU capacity planned for that year or later years.

Oracle also announced expanded work with OpenAI at the event. The Register reported plans for ChatGPT integrations with Oracle applications and cloud services, alongside a separate Oracle product using OpenAI models. Those announcements extend Oracle’s AI partnerships as it prepares to add Google’s chips to its infrastructure.

Google splits its eighth-generation TPU range

Oracle says the eighth-generation range has two architectures: TPU 8t for training and TPU 8i for inference. It claims TPU 8t provides nearly 3x the compute performance per pod of the preceding generation. Oracle describes TPU 8i as designed for the repeated processing demands of specialised AI agents.

Google says its eighth-generation TPUs deliver up to 12.6 PFLOPS of native FP4, or four-bit floating-point, compute per chip, according to a Google Cloud post about the partnership. Google also claims up to 2x the performance per watt of the previous generation. Those are Google’s hardware performance claims, rather than results reported for TPU workloads running on OCI.

For large training runs, Google says its systems sustain more than 97% goodput across thousands of chips. Goodput is the share of a cluster’s time spent on productive computation. Google attributes that result in part to networking that can route around hardware or connection faults without restarting a job.

Google says the latest TPUs can deliver up to 80% better performance per dollar for low-latency serving. Its post links that claim to changes in on-chip memory, high-bandwidth memory capacity and the connections between chips. The measure concerns Google’s TPU systems, rather than a published price for using the chips through OCI.

TorchTPU gives Google a PyTorch route

Google says its TPUs train and serve Gemini models and that TorchTPU, its native PyTorch stack for the chips, is available as open-source software. AI Affairs previously reported Google’s release of TorchTPU a day after Amazon released a rival tool for PyTorch users.

Amin Vahdat, Google’s senior vice president and chief technologist for AI and Infrastructure, said offering TPUs through OCI would extend access to organisations building training and inference workloads. Google Cloud says customers will be able to use the chips on either its own platform or Oracle’s once the OCI deployment begins.

Oracle says TPU capacity on OCI will become available in calendar year 2027 and increase over a multi-year period.