Neuron Fit: will my PyTorch job run on Trainium, and what will it cost?

I found no product that combines the three parts of this idea: run your repo through TorchNeuron, list what breaks, and report cost per step or per million tokens against H100 rental prices. AWS has the pieces in separate places. There is an analyze tool tied to the deprecated XLA path, Transform agents that port kernels, and benchmark pages that only cover training on Trn1. Optimum Neuron lists validated HF models but publishes no throughput or cost comparison that gets re-run with each release. The timing also helps. PyTorch/XLA training support ended with Neuron 2.30, so every existing Trainium training user has to re-check their models on TorchNeuron. That makes a public board re-run on each beta release useful right now. The risks are real, though. TorchNeuron is still a private beta that you get through an AWS account manager. Publishing beta results on a public board may break beta terms, and you need access before you can do anything. AWS is also clearly moving into this tooling itself with the Transform agent, the analyze tool and benchmark docs, and it could ship a free fit checker. The paying customer is also awkward. Startups on Activate credits are spending AWS's money, so their pain is wasted engineering time, not compute cost, and they may be unwilling to pay cash for a report. The likeliest way to win is to run the public board as a neutral, independent source for the GPU cost comparison, which AWS won't publish credibly. That could lead to paid fit reports, or to sponsorship from AWS or partners. A standalone SaaS sold to seed-stage startups looks weaker.

• AWS Neuron analyze tool (neuron_parallel_compile / analyze for training)

  AWS's own checker gives a support percentage and flags unsupported ops. It is closest to the 'what breaks' half of the idea. It counts XLA/HLO instructions rather than aten ops, it is built for the old XLA path, which has now been archived for training, and it has no cost comparison.

• AWS Transform custom agent for Trainium kernel migration

  Announced July 2026. It converts PyTorch modules and Triton kernels to NKI code, then compiles and benchmarks them on real Trainium before opening a PR. It runs from a CLI or MCP and covers the 'suggested workarounds' part. It is free first-party tooling, so it is the biggest overlap risk.

• Hugging Face Optimum Neuron

  Keeps an official list of validated models and tasks for Trainium and Inferentia, plus a public compile cache. It partly overlaps with a public pass/fail board for HF models, but it publishes no regularly re-run throughput or cost table.

• AWS Neuron benchmarks pages

  Official throughput numbers with reproducible configs. Training coverage is only on Trn1, there are no Trn2 or TorchNeuron numbers and no GPU cost comparison.

• AWS re:Post guide: is my model supported on Neuron

  A manual guide to checking compatibility. Its existence shows the question is real and that no tool answers it yet.

• Yotta Labs Trainium diffusion benchmarks

  A one-off third-party study reporting up to 4.1x throughput per dollar against H100 on diffusion models. This is the kind of cost-per-dollar comparison the product would automate, but it is not a tool.

• GPUPerHour Trainium vs GPU guide

  An editorial guide that recommends a pilot before switching and counting porting cost in total job cost. It is content, not a product.

Checked 2026-10-20