# SemiAnalysis tests Trainium3 PyTorch stack, finds better profiling than Nvidia

Digest AI · Hardware & Compute · published 2026-10-23T19:06:44Z

Canonical: https://digestai.news/story/semianalysis-tests-trainium3-pytorch-stack-finds-better-profiling-than

## Summary

SemiAnalysis engineers spent weeks testing Amazon's Trainium3 racks with the new TorchNeuron backend and report a native PyTorch experience that works in eager mode with torch.compile and standard distributed APIs. They trained Qwen3-32B on 16 Trainium2 chips at 23.89% MFU (eager) and 30.47% MFU (compiled), and tested gpt-oss 20B on Trainium3 with a 27-minute first-step compilation time. Trainium3 beat Nvidia GB300 and B200 on matrix-multiplication MFU up to 100 GFLOP and again past 1 TFLOP. FlexAttention passed 89 of 96 tests; the seven failures were compilation-time budget issues at large batch and sequence length. AWS runs automated quality control on up to 80,000 physical cores against 100,000+ Hugging Face models, checking numerics, performance and CPU fallbacks. The Neuron Explorer profiler drew praise from Anthropic's performance lead and SemiAnalysis for its timeline view, dependency visualization and intra-kernel detail, which they say exceeds Nvidia Nsight and AMD tooling. Bugs remain and the compiler is slow, but the trajectory is rapid, aided by agentic coding.

## Key points

- Trainium3 TorchNeuron now runs native PyTorch eager mode with torch.compile
- Qwen3-32B hit 30.47% MFU compiled on 16 Trainium2 chips; gpt-oss 20B compiled in 27 minutes
- Neuron Explorer profiler rated better than Nvidia and AMD tools by Anthropic and SemiAnalysis

## Why it matters

A credible non-Nvidia training stack with strong profiling could give researchers and Tier-2 labs a real alternative for model development, not just inference.

## Sources

1. [Amazon’s Trainium Takes a Strike at the Training King](https://newsletter.semianalysis.com/p/amazons-trainium-takes-a-strike-at) (SemiAnalysis, 2026-10-23)

Part of the developing story: [Amazon Builds Custom AI Silicon Ecosystem](https://digestai.news/thread/amazon-plans-to-sell-8-billion-of-nvidia-chips-to-investors) (6 stories)

## Cite

Digest AI, "SemiAnalysis tests Trainium3 PyTorch stack, finds better profiling than Nvidia", 23 October 2026, https://digestai.news/story/semianalysis-tests-trainium3-pytorch-stack-finds-better-profiling-than

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/semianalysis-tests-trainium3-pytorch-stack-finds-better-profiling-than.json