Tech
Google officially details TorchTPU as an open-source PyTorch backend for TPUs, with Meta collaboration, one day after Amazon's TorchNeuron.
The Stack · 2026-10-20 · Source ↗
Google officially details TorchTPU as an open-source PyTorch backend for TPUs, with Meta collaboration, one day after Amazon's TorchNeuron.
The Stack · 2026-10-20 · Source ↗
Google officially detailed TorchTPU, the native PyTorch backend for its Tensor Processing Units, at the PyTorch Conference week in San Francisco. The project was first reported inside Google in December 2025 and had been running under an internal codename. The Tuesday session confirmed what Google and Meta engineers had hinted at for months: TorchTPU ships as an open-source PyTorch extension that registers TPUs through the framework's custom-device slot, so standard PyTorch APIs run on a device named "tpu" without model code changes.
What ships
Eager execution runs natively on the TPU without requiring the PyTorch/XLA compilation flow that users have struggled with for several years. torch.compile is supported in production and tracks the standard FX graph capture path, with the Neuron-equivalent lowering through Torch-MLIR into OpenXLA's StableHLO contract. Distributed APIs including DDP, FSDP2 and tensor parallelism work over the TPU inter-chip interconnect. Google published documentation covering TPU v5p, Trillium, and the newer Ironwood generation that ships inside AI Hypercomputer. Meta engineers contributed on the FSDP2 side and ran the sweep against production transformer stacks.
"The TPU should be an obvious choice for any PyTorch user to target. Getting access through PyTorch has always been difficult. We are changing that this year." — Lee Howes, engineering lead on TorchTPU, Google
Where it sits in the stack
The release lands one day after Amazon's TorchNeuron, which registered Trainium through the PrivateUse1 extension point and shipped under Apache 2.0 for the broader PyTorch Conference keynote. Google's engineering team framed TorchTPU as the companion piece: both backends target the same PyTorch eager and compile APIs, both ship as open-source projects, and both aim to lower the cost of moving models off CUDA. A Google Cloud spokesperson described "massive, accelerating demand" for TPU and GPU infrastructure and said the company wants developers to have flexibility "regardless of the hardware they choose." Thomas Kurian, Google Cloud CEO, cited TPU "price-performance and efficiency" in the context of Anthropic's expanded TPU use.
What remains open
Google did not disclose a performance comparison against CUDA on identical workloads. Repository access, exact license tier beyond "open source," and sample throughput numbers on Trillium versus H200 are expected in a follow-up developer blog post. The release does not retire PyTorch/XLA; existing XLA users remain supported, and TorchTPU is positioned as the eager-first entry point that avoids the XLA-only ergonomics.