Google's AI Chips, TPUs, Begin Being Used in Meta Data Centers
Google announced that its custom artificial intelligence chips, TPUs, are now being used in Meta's data centers. Thanks to the new PyTorch-based SGLang and vLLM inference engines, AI workloads can be run on the hardware without requiring code modifications.
Google announced that its self-developed artificial intelligence accelerators, Tensor Processing Units (TPUs), have started being actively used in Meta's data centers.
TPUs developed by Google have been integrated into Meta's data centers to support artificial intelligence production workloads. This integration was made possible thanks to the new PyTorch-based SGLang inference engine announced in Google's developer blog. Additionally, a TPU-compatible open-source version of vLLM, another widely used inference engine, has been released.
Previously, these inference engines used Google's own JAX artificial intelligence framework to run on TPUs. With the new updates, the hardware has become a native option for PyTorch, allowing the system to run without users needing to make major code changes.
Sources
- tomshardware.com · Link