Amazon released TorchNeuron, which moves a PyTorch application onto its Trainium chips with a one-line configuration change and no code rewrite, while it contributes changes upstream that remove chip-specific code from PyTorch. Hugging Face says 89% of the 100 most-used text-generation architectures now run on Trainium unmodified, up from 46%, ahead of a public beta that starts October 20; it does not describe how that was measured. Amazon says about 85% of AI applications are written in PyTorch, so porting effort is the barrier it is now addressing, and the only customer result it cites is Deepgram processing 70% more audio per dollar. Amazon shares closed up 5.0% at $262.45 on the day. (TechCrunch)
Public Markets
Micron rose as the Philadelphia Semiconductor Index gained about 2.7%, ending an eight-session losing streak. Memory led, with SanDisk up 5.5% and SK hynix ADRs up 4.7%. (Stock Analysis)
Meta$705.44▼ 0.5%Mkt Cap: $1.8T
Meta said phase one of its Prometheus cluster in Ohio is complete and trained its Muse models, with goodput above 85% for pre-training and a 99% drop in interruptions from the mid-2025 peak. Synchronized training swings facility load by 30% to 40%, so power transients now rank beside FLOPS as design constraints. (Engineering at Meta)
Emerging
Decode inside the HBM stack: A paper from Sanghyeok Han, Inseong Choi, Byungkuk Yoon, Gyeonghwan Park and Jae-Joon Kim, A Hybrid GPU and Processing-Near-Memory Architecture with Shared DRAM for Efficient Prefill-Decode Disaggregation, adds processing-near-memory units to the HBM logic die that run the full decode phase while the GPU runs prefill, sharing bandwidth over a common HBM interface. It reports up to 2.23x end-to-end throughput over conventional prefill-decode disaggregation on server-scale LLMs, the abstract's only quantitative result and one from the authors' own evaluation. The design moves part of the inference workload into the memory stack, which ties its payoff to what HBM vendors choose to build into their base dies. (arXiv)