NVIDIA AI Updates: September 15, 2026

1. Transformer Engine Pushed Dropless MoE Training in JAX to a 10.4x Throughput Gain

NVIDIA. Dropless MoE processes every token without dropping or padding when experts receive uneven loads, which preserves training quality but produces ragged tensor shapes that standard kernels handle badly. Transformer Engine addresses this with a grouped GEMM that runs all expert matmuls in one kernel using actual token counts, eliminating device-to-host transfers and worst-case padding, plus expert parallelism over an NCCL EP backend that fuses dispatch and combine and deduplicates tokens crossing the network. On DeepSeek-V3 671B running on GB200, end-to-end throughput went from 103 to 1,068 TFLOPS per GPU, with 97% scaling efficiency at 1,024 GPUs on GB300 NVL72. Source

2. Perplexity’s Local Agent Shipped for Windows on RTX GPUs With 24 GB or More

NVIDIA. Portable Computer, the local version of Perplexity’s Computer agent, is now available in the Perplexity Windows app on GeForce RTX and RTX PRO systems with at least 24 GB of VRAM, with DGX Station support listed as coming soon. It ships with Qwen 3.8 27B, post-trained for Perplexity Computer and optimized for RTX GPUs, and includes a built-in browser and sandbox so multistep tasks run without sending data off the machine. The Windows release follows existing Linux and DGX Spark support. Source