Skip to content
Stage · Medium2 outlets · 2 articlesTech

Amazon's TorchNeuron runs PyTorch on Trainium with one change, but 22,000 tests are skipped

Amazon's TorchNeuron runs PyTorch on Trainium with one change, but 22,000 tests are skipped — Newsmesis editorial illustration
— Newsmesis illustration

What happened

Amazon's one-line switch from standard chips to its Trainium AI chips ships with about 22,000 PyTorch tests skipped, according to a PyTorch blog post by Amazon engineers cited by The Next Web. The tool, TorchNeuron, is an open-source backend (Apache 2.0 licence) that lets existing PyTorch code run on Trainium2 and Trainium3 by changing a device setting. It enters public beta on 20 October at the PyTorch Conference in San Jose. GeekWire frames it as an attempt to make leaving Nvidia hardware about as simple as changing a setting, but the first release lacks some PyTorch features, such as dynamic input shapes.

What we know

The Next Webthenextweb.comGeekWiregeekwire.com
  • TorchNeuron goes into public beta on 20 October at the PyTorch Conference in San Jose and is released as open source under the Apache 2.0 licence, per The Next Web; GeekWire also reports a public beta with open-source release.
  • It supports Trainium2 and Trainium3 and plugs into PyTorch's built-in mechanism for adding new hardware, so standard testing and scaling tools still work, per The Next Web.
  • The first release does not support some PyTorch features, including dynamic input shapes, and about 22,000 PyTorch tests are skipped, per The Next Web citing the Amazon engineers' PyTorch blog post.
  • Early users named by The Next Web include voice AI startups Cartesia and Deepgram and the Lieber Institute for Brain Development; Hugging Face has added Trainium support to its AI tools.
  • Amazon said it ran the same training script on a standard chip and on Trainium and that the model learned at the same rate on both, per The Next Web. This is Amazon's own demo result.

What we don't know yet

  • Whether the 'same learning rate' demo result holds beyond one Amazon-run training script; no independent benchmark is cited in the sources.
  • How much real-world code the roughly 22,000 skipped tests and missing features (such as dynamic shapes) will block from running unchanged.
  • How the tool affects Trainium adoption. The Next Web cites Amazon saying its custom chip business has passed a $25 billion annual revenue run rate, a company figure not independently verified here.
  • Whether Amazon will acquire Decart, an AI chip software startup and TorchNeuron launch partner that Amazon was reported to be in talks to buy; The Next Web gives no source for that report.

Why it matters

Months of code rewriting have been a main cost of moving AI workloads off Nvidia's CUDA-based software stack. If TorchNeuron works as described on real workloads, AWS customers training in PyTorch can test Trainium with far less engineering effort. The skipped tests and missing features show where that promise currently stops.

Claims

Related coverage

Microsoft's Maia 300: 96 chips per rack, 40% more tokens per megawatt than GPUs

Tech ·

Microsoft's Maia 300: 40% more tokens per megawatt, but not until 2027

Tech ·

Nvidia's RTX Spark laptops ship from six makers, with up to 128GB of memory

Tech ·

Why you can trust this story

67%

Source map · 2 outlets / 2 articles

How this credibility score is calculated

  • Source reliability
  • Corroboration
  • Primary source
  • Atom-verified claims
  • No contradiction
  • Settled
  • Claim attribution
  • AI disclosure

Atom-verified claims matched cited source text. Weights are fixed and explainable.

Human accountability

  • 2 independent origins / 2 sources
  • Drafted by the Newsmesis judgment agent (agent-cli); human approval required before publishing.

Verification ledger · 9

  • Claim: TorchNeuron enters public beta on 20 October at the PyTorch Conference in San Jose.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: TorchNeuron is open source under the Apache 2.0 licence.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: TorchNeuron supports Trainium2 and Trainium3 chips.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: The first release does not yet support some PyTorch features, such as dynamic input shapes.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: About 22,000 PyTorch tests are skipped in the first release, according to the PyTorch blog post by Amazon engineers.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: Amazon says developers can target Trainium by changing a single device setting in existing PyTorch code.
    Status: ConfirmedVerification trailVerifiedChecked by verification engine

    At least one extracted atom matched cited source full text.

  • Claim: In an Amazon demo, a model learned at the same rate on Trainium as on a standard chip.
    Status: UnconfirmedVerification trailUnverifiedChecked by verification engine

    No atom-level source-text verification is available.

  • Claim: Amazon says its custom chip business has passed a $25 billion annual revenue run rate.
    Status: UnconfirmedVerification trailUnverifiedChecked by verification engine

    No atom-level source-text verification is available.

  • Claim: Amazon was reported to be in talks to buy Decart, a TorchNeuron launch partner.
    Status: UnconfirmedVerification trailUnverifiedChecked by verification engine

    No atom-level source-text verification is available.

Update log · 0

    AI accelerates. High-risk topics wait for a human editor.