Amazon and Google run PyTorch Nvidia code on their chips
In brief
- Amazon and Google launch TorchNeuron and TorchTPU to run PyTorch Nvidia code on their chips with minimal changes.
- CUDA/Triton kernels are unsupported: rewriting is required using NKI, Pallas or Tokamax.
- Deepgram reports a 11 to 70 % efficiency gain on Trainium2 after months of porting.
On October 20, 2026, Amazon and Google released TorchNeuron and TorchTPU as open source tools to run PyTorch code designed for Nvidia GPUs on their own chips, Trainium and TPU. Both companies claim that minimal configuration changes or code adjustments are enough to migrate existing models.
TorchNeuron and TorchTPU replace older PyTorch/XLA gateways. They target Trn2/Trn3 instances and TPU v5e to v8t respectively. CUDA or Triton kernels are not supported: they must be rewritten using specific tools (NKI, Pallas, Tokamax). Deepgram, the only client to detail its porting, reports a 11 to 70 % gain in audio processed per dollar on Trainium2 after months of kernel rewriting and a shift to a Rust-based stack.
These solutions are still under development: TorchNeuron requires PyTorch 2.14.0 and excludes dynamic shapes, while TorchTPU marks some features as experimental. AMD and Intel plan to merge their PyTorch optimizations for x86 into oneDNN, though no timeline is given. No independent measurements validate the reported gains.
What it changes for you
- Industry & ManufacturingManufacturers using PyTorch for embedded or cloud AI can cut costs by migrating to Trainium or TPU, provided they rewrite specific kernels.
This summary gives you the essentials. The full article, its figures and quotes are on ActuIA.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error