Desk
NVIDIA hardware, software, and developer news.
NVIDIA’s DGX Spark adds a 64GB Grace Blackwell desktop — 1 petaFLOP FP4, clusterable to 128GB — aimed at local agents that would otherwise burn metered API tokens all day.
Nvidia’s 2019 Shield TV Pro jumps $100 to $299.99 because memory got dear — a hobbyist streaming box now paying the AI boom’s hardware tax.
A cheaper Grace Blackwell personal box with partner SKUs and two-node Sync clustering — NVIDIA is selling local agents as a starting point, not a cloud apology.
Astra Ultrafast on Blackwell promises up to 8x token speed for agent loops — OpenAI using its own models to tune inference is the quiet industrial story.
NVIDIA’s DOCA agent skills turn BlueField coding agents from guessers into checklist passers — 19% to 100% on a 65-prompt eval.
DIN Deploy open-sources C++ ONNX Runtime + TensorRT RTX samples — Whisper to FLUX.2 — aiming at real local apps, not notebook demos.
NVIDIA’s AI-factory pitch: megawatt-scale plants (~$60M each) win on productive tokens-per-MW, multi-year durable GPUs, and fungible workloads — citing SemiAnalysis AgentX gains for Vera Rubin.
NVIDIA engineers cut Najdi/Hijazi ASR word error from 55% to ~30% by fine-tuning Nemotron 3.5 — without tanking other Arabic dialects or English.
NVIDIA’s recsys-examples path serves Hierarchical Sequential Transduction Unit ranking models via Dynamo-Triton — up to ~4.5x speedups on RTX PRO 6000 Blackwell in cited tests.
cuObject client/server libraries hit GA for RDMA object storage, SCADA Server SDK lands for GPU-initiated I/O, and xio-sig courts Google Cloud and Microsoft for interoperability.
NVIDIA’s 26th Graduate Fellowship cycle opens worldwide — up to $60,000, Ph.D. year-one-plus required, mandatory summer 2027 research internship, applications due October 30, 2026.
NeMo Relay gives Hermes Agent ordered lifecycle traces and trajectories — so you can tell whether a harness change fixed tasks or just burned more tokens.