Trainium3 Tested: 30.47% MFU on Qwen3-32B, but GPT-OSS Compiles 27 Minutes Before Step One
demian_ai · x · 2026-11-03
A reshared SemiAnalysis analysis of Amazon's Trainium chip: the hard part isn't silicon but software—compiling PyTorch code (which grew up on Nvidia GPUs) to custom hardware. In real training tests with AWS's new native PyTorch support, Qwen3-32B hit a respectable 30.47% MFU, but compiling GPT-OSS took 27 minutes 17 seconds before the first training step. Amazon's compiler was built inference-first; Trainium has shone at inference (Anthropic runs production inference and RL rollouts on Trainium2/3), and AWS's shift to first-class eager-mode PyTorch with torch.compile is the right direction, though maturity still lags CUDA badly.
More from Infra
- Mistral ML4 goes open-weight: trained from scratch on just 10MW, claims top open model in code agents — _AndrewZhao · 2026-11-03
- Databricks: Agents now run AI research loops, training jobs grew 8x in a year — matei_zaharia · 2026-11-03
- RTX 5080 reportedly discontinued, store prices nearing $2,000 — PAstynome · 2026-11-03
- Inference startups hitting $100M+ ARR are just reselling prebooked compute, not real businesses — abhiadesai · 2026-11-03
- Stripe Data Pipeline adds real-time sync to Lakebase Postgres with seconds-level updates — matei_zaharia · 2026-11-03
- AI datacenter capex hits $1.2T in 2026 as growth peaks at 56% — the 'second derivative' worry explained — TiernanRayTech · 2026-11-03