Desk
NVIDIA hardware, software, and developer news.
From Blackwell Ultra to Vera Rubin, Nvidia says HBM per GPU rose about 3% while agentic throughput per dollar rose up to 30x on SemiAnalysis AgentX. Its answer is a memory hierarchy, compact formats and KV-cache offload. The direction is right; the numbers need outside checks.
NVIDIA's developer deep dive names five co-design levers — NVFP4, Wide-EP, Dynamo, Vera CPU, and CMX — that push Vera Rubin to up to 30× GB300 throughput per megawatt on the same 288 GB HBM.
Nvidia’s cloud service is shipping DLSS 5 with 3D-Guided Neural Rendering in Hogwarts Legacy, tuned by Avalanche to keep the art style. It shows how streaming can deliver next-gen graphics without a new GPU, though the full experience sits behind the Ultimate tier.
Box Content Retrieval embeds Nvidia’s open Nemotron models in its core so agents get relevant chunks instead of reading whole files. Nvidia expects 4x speed, 20% fewer errors and 90% fewer tokens. The moat isn’t the open weights; it’s Box’s permissions and metadata.
Nvidia is running its own DGX SuperPOD ops with a chat of specialized agents that correlate compute, network, hardware, and scheduling failures — on-prem, with Nemotron, and read-only until SREs trust it.
CUDA 13.5 takes Vera Rubin from early access to production GA and piles on MPS scheduling, multi-GPU Tile communication, Python ergonomics, and Nsight tools — the software floor for the next Nvidia hardware generation.
NVIDIA is pitching a disciplined loop for fixing agents: read the traces, write a test for the failure, change the harness, and prove it on tasks the fixer never saw.
NVIDIA's pitch for the AI-defined fab is mostly partner speedups, and the useful parts are real, but this is a vendor describing a market it sells into.
NVIDIA is pitching cities a shared AI stack for services, planning and surveillance cameras. The service gains are worth having, but turning camera networks into AI agents needs public rules first.
NVIDIA open-sourced the tool it uses to figure out why giant training jobs die, and its best feature is admitting when the evidence isn't enough to say.
NVIDIA wants the 3 a.m. restart of a failed training job to stop being a human's job, and it ships with a sensible default: watch first, act later.
NVIDIA wants Vera Rubin racks running Kubernetes from the first day, and its tested recipe approach is the right fix for a messy bring-up, even though inference is not ready yet.