PyTorch opens to TPUs and Trainium without sidelining Nvidia
Google and AWS add native backends to PyTorch for their TPU and Trainium chips. The goal: make their accelerators more accessible, without promising effortless maximum performance.
At the PyTorch Conference North America 2026, Google and AWS presented two native backends for PyTorch: TorchTPU for Google's TPUs and TorchNeuron for AWS's Trainium chips. The idea is simple but strategic: to allow developers to more easily move their workloads outside the Nvidia ecosystem, which remains dominant. In practice, portability is improving, though the promise of “plug-and-play” performance is not yet fulfilled.
Deux backends pour rendre les puces maison plus attrayantes
TorchTPU is now available on GitHub and targets Google's TPUs, while TorchNeuron has entered public beta for Trainium2 and Trainium3 at AWS. Both tools aim to make PyTorch a more natural entry point to accelerators designed internally by the two groups. According to Google's text, TorchTPU allows running existing PyTorch workloads without JAX dependency in the runtime, with code changes described as “trivial code changes.”
AWS pushes the same logic with TorchNeuron. Switching to Trainium can, according to the company, often be limited to a simple change to .to('neuron') for eager execution, then selecting the Neuron backend for compiled models. This approach matters because it removes some friction that often discourages teams from testing chips other than Nvidia’s. But it does not eliminate the need to optimize for the chosen target.
Le vrai sujet reste la performance, pas l’annonce
Both players know the issue is not just about facade compatibility. For distributed workloads, PyTorch indicates that TorchTPU’s default approach, called “Skinny Collectives,” can cause a performance penalty of about 30 to 40% compared to an optimized approach, because it prevents the XLA compiler from overlapping computation and communication. In other words, portability is there, perfect fluidity is not.
PyTorch recommends, for production workloads aiming for the best throughput, additional adjustments and more specialized stacks, such as Pallas kernels, vLLM-TorchTPU, SGLang-TorchTPU, or TorchTitan-TorchTPU. On the AWS side, the company specifies that most teams can try Trainium without full rewrite but implies some workloads will require more changes. One case cited in the documentation required three changes to adapt the Laplace 3B model: using native PyTorch for the GLA attention kernel, redesigning Muon communication, and compiling the sequential backward pass.
PyTorch devient aussi une bataille d’infrastructure
The movement goes beyond just chip choice. According to Amazon, 85% of AI researchers and developers write their code in PyTorch, making it a decisive software layer for hardware vendors. Chris Lattner, Qualcomm’s Executive Vice President of AI, summarized the challenge by saying: “We all know that PyTorch has done an amazing job at unifying the modeling layer, making it easy to build and train models for a wide variety of different applications, but we still have the bottom part of the stack that the hardware vendors have to make for themselves. This includes the kernels, the compilers, and low-level system software.”
In this context, Google, AWS, Qualcomm, and others seek to fill this “bottom of the stack” which includes kernels, compilers, and distributed training tools. Google also announced integrations with Hugging Face, vLLM, SGLang, llm-d, TorchTitan, Ray, and TorchRec, as well as the arrival of LiteRT in the PyTorch ecosystem. The message is clear: if the model attracts attention, the infrastructure running it always ends up taking over.
La montée en charge oblige PyTorch à se réorganiser
PyTorch itself is absorbing this growth. Alban Desmaison, co-maintainer of the project, told the audience that more than 600,000 GitHub projects rely on PyTorch and about 150 pull requests arrive daily. Mark Collier, Executive Director of the PyTorch Foundation, announced a 60% increase in active contributors year-over-year, 30% more participating organizations, and 40% more pull requests. These figures indicate less a victory than a tension: the framework’s popularity strains testing, infrastructure, and release operations.
Edward Yang, also co-maintainer, put the debate at its most concrete level: “At the end of the day, what you care about is how long it takes for you to get the final checkpoint that you’re actually going to serve.” Teams want smaller, simpler-to-profile blocks that are easier to fix when something breaks on the critical path. The obsession with pure benchmarking yields ground to a more prosaic need: shipping a usable model without spending nights debugging opaque layers.
Les usages business rappellent pourquoi ce socle compte
Shopify showed at the conference how PyTorch also serves well-defined business tasks. The company described a continuous learning flow based on a reference dataset, an evaluator calibrated on human judgments, then a loop where production failures feed retraining. For its customer profiling application, Shopify started with about 29,000 examples and enriched its model at each iteration. “It’s important to keep in mind that human agreement is the ceiling,” explained Cody Mazza-Anthony, ML engineer at Shopify. “We’re not looking for perfect here. We’re looking for as good as human.”
This type of use shows why PyTorch’s openness to other chips really matters. If companies can more easily move their workloads between Nvidia, Google, and AWS, they gain negotiating power, some cost margin, and more technical freedom. Still, openness does not erase trade-offs: you can simplify porting, but not abolish the laws of physics, unfortunate for brochures.
Key points
- TorchTPU and TorchNeuron target existing PyTorch workloads.
- Google and AWS publish their backends on GitHub.
- PyTorch has over 600,000 GitHub projects.
- Alban Desmaison cites about 150 pull requests per day.
- Shopify started with 29,000 examples for a business case.
- According to PyTorch, TorchTPU can lose 30 to 40% performance in distributed mode.
By the numbers
- 85% — of AI researchers and developers write in PyTorch, according to Amazon at the 2026 conference.
- 600,000+ — GitHub projects built on PyTorch, according to Alban Desmaison.
- 150 — daily pull requests received by PyTorch, according to Alban Desmaison.
- 60% — annual increase in active contributors, according to Mark Collier.
- 30 to 40% — potential performance penalty on TorchTPU in distributed mode, according to PyTorch.