SemiAnalysis tests Trainium3 PyTorch stack, finds better profiling than Nvidia
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.…
1 source
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
The story so far
6 episodes →- SemiAnalysis tests Trainium3 PyTorch stack, finds better profiling than Nvidiathis story
Read the original at SemiAnalysis · by Kimbo ChenOpen source ↗
Topics · follow one to build your own front page
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Nvidia RTX 5090 shipment $470,000 stolen from Brazil Airport, card listed $6,000 online · 3 src
- HPE plans first Moon supercomputer benchmarks with Spaceborne Computer-4 in late 2026 · 2 src
- Nvidia CEO Jensen Huang unveils DSX POD to cut AI data center deployment from months to 14 days · 21 src
- Nokia and Nvidia target 50% spectrum efficiency by 2027 · 6 src
- OpenAI scales Azure Cosmos DB to serve 1 billion ChatGPT users · 1 src
Comments
via GitHub Discussions