NVIDIA tutorial: reading data into GPU memory via GPUDirect Storage in Python

Databricks and SAP announced a joint push to connect data and AI capabilities end to end for enterprise AI transformation. The post argues capable models still lack the business context needed to deliver value.

The trtedge CLI downloads a model and builds TensorRT engines for Spark's new 64GB configuration, tested with TensorRT Edge-LLM 0.11.1 and TensorRT 10.16.1.11. DSpark speculative decoding speeds generation, and the same runtime deploys to Jetson Thor, DRIVE AGX Thor and IGX Thor.

A single X post by sgondala2, shared on Hacker News, sketches how to structure an engineering organization around AI. No concrete metrics, tooling, or case data appear in the available snippet.

NVIDIA Halos SAFE AV connects real-world testing, simulation, and vehicle hardware into one continuous evaluation loop. Omniverse NuRec reconstructs real driving scenarios while Cosmos generates variations across traffic, weather, and edge cases.

A merged vLLM pull request adds PLE disk offload, landing in nightly builds rather than a tagged release. The change lets PLE data be paged to disk instead of held in memory.