partnership

Anyscale, Google Cloud add native Ray support for PyTorch on TPUs

0Importance
closedStatus
2News media
Published

Anyscale and Google Cloud said Ray now runs PyTorch natively on Google TPUs on GKE, with one worker assigned to each TPU chip and TPU-aware scheduling. Google also open-sourced TorchTPU, a library for native PyTorch support on its TPUs.

Key claims & sources

✓
Anyscale and Google Cloud said Ray now supports PyTorch on Google TPUs natively on GKE.
✓
The setup uses one worker per TPU chip and TPU-aware scheduling.
✓
Google open-sourced TorchTPU to bring native PyTorch support to its TPUs.
✓
The TorchTPU release was presented as a way to improve PyTorch integration with TPU hardware and raise efficiency.

What it means

PyTorch is one of the main machine learning frameworks, and Google TPUs are specialized chips used for AI workloads on Google Cloud. Ray is the distributed computing layer that helps teams spread training and inference across hardware, and GKE is Google Kubernetes Engine, the managed container platform where these workloads run. The addition of native TPU support for PyTorch on Ray, alongside TorchTPU being open-sourced, points to a push to make Google’s hardware easier to use from a widely adopted AI stack.

For teams building on PyTorch, these releases narrow the amount of custom plumbing needed to run on TPUs. One worker per TPU chip and TPU-aware scheduling make the hardware layout more explicit inside Ray, which can simplify deployment decisions for engineers already using Google Cloud. The practical change is that developers working in the PyTorch ecosystem have a clearer path to target TPUs without leaving their existing framework.

The larger trade-off is between open, portable developer tooling and the pull of a tightly integrated cloud stack. Native support inside Google’s own environment can reduce friction, but it can also deepen dependence on Google Cloud and its TPU ecosystem for teams that adopt it. At the same time, opening TorchTPU suggests a bid to broaden access to TPU support rather than keep it locked inside a closed implementation.

Still open is how quickly developers outside Google’s own environment adopt TorchTPU and Ray’s TPU support in production. The releases do not show how many teams will switch, which workloads fit best, or whether these tools will become the default path for PyTorch on TPUs. It also remains to be seen how much of the surrounding ecosystem, from orchestration to debugging and monitoring, is fully covered by the new support.

Market reaction i

Total crypto marketall crypto assets combined no single asset in the event: read against the whole market

The total crypto market rose 0.3% in the 4 hours after the first headline, after falling 0.3% in the 2 hours before.

$2.8T market cap at the headline
-0.3% 2 h before
+0.3% 4 h after
+0.6% 12 h after
First headline

47 prices, one every 30 minutes

Coverage timeline

Every piece in order, with the time its news medium declares — if it is not believable we use ours and say so — and what it added to what was already known at that point. Arriving first is not the same as adding something.

first
Oct 20 16:26
+3 h 9 m
Oct 20 19:35
Crypto Briefing context
Its angle: development, integration, efficiency