NVIDIA AI Updates: October 31, 2026

1. Breakthrough Listen Puts AI Agents on the Allen Telescope Array With Holoscan and RTX PRO 6000 GPUs

NVIDIA. Breakthrough Listen and the SETI Institute have added AI agents to their technosignature search at the Allen Telescope Array in Northern California. From a single natural-language request, agents steer the antennas, calibrate and focus the instrument, detect and repair data errors, rewrite signal-processing blocks and plot results. Nexus, a web-based observatory orchestrator, serves as the agent harness and the single source of truth for instrument state, compute resources and pipelines. The data path runs on NVIDIA Holoscan: the DAQIRI library moves sensor data straight into GPU memory during observation, the SETI Institute’s CyberEther framework does the signal processing on cuFFT and cuBLAS, and inference runs on premises on RTX PRO 6000 Blackwell GPUs at up to 2.4 terabits per second. The pipeline produced the first real-time AI detection of a pulsar in the Crab Nebula. The team is evaluating open models including Nemotron for cases where data locality and turnaround time matter most. The post gives no accuracy figures and does not say whether Nexus or DAQIRI are open source. Source

2. gds-diag Open-Source CLI Audits GPUDirect Storage Setups and Ships With an Agent Skill

NVIDIA introduced gds-diag, an open-source command-line utility that assesses a system’s GPUDirect Storage environment, identifies configuration problems and suggests fixes. A config-audit subcommand reviews GDS configuration files and environment variables, and an optional agent skill lets coding agents run the same diagnostics from natural-language prompts. Both are distributed through the NVIDIA Magnum IO GitHub repository; no version number is given. GDS moves data between storage and GPU memory without a CPU bounce buffer, and it depends on driver, filesystem and kernel settings that are easy to get wrong, so a scripted checker is useful for anyone standing up training or inference nodes. Source

3. AI Cluster Runtime Recipe Pins Dynamo Platform 1.4.2 and Validates It on GB200 and GB300 Clusters

NVIDIA published a deployment and validation walkthrough for a Dynamo inference runtime on Kubernetes using NVIDIA AI Cluster Runtime (AICR), the open-source tool that snapshots a cluster, generates a recipe, renders a Helmfile or GitOps bundle (Argo CD and Flux are supported) and runs conformance checks. The AICR v1.0.0 recipe for Amazon EKS on GB200 NVL72 and GB300 NVL72 pins Dynamo Platform 1.4.2 alongside GPU Operator v26.7.1, KAI Scheduler v0.16.9, Grove v0.1.0-alpha.13, NVSentinel v1.25.0, NodeWright v0.19.0, DRA GPU Driver 0.5.0, Agent Gateway v1.5.0 and Kube Prometheus Stack 84.4.0. Dynamo 1.4 moves to the nvidia.com/v1beta1 DynamoGraphDeployment CRD, replacing v1alpha1, and uses Kubernetes resources for service discovery, so NATS is now opt-in; requests travel over TCP and KV-cache events over ZMQ. Two routing paths are documented: the Dynamo Frontend in KV router mode, which needs no gateway, and Gateway API Inference Extension with a Dynamo endpoint picker that embeds the Rust KV router through CGO. Dynamo 1.5 is planned to replace that with a native Rust EPP. aicr validate covers cluster baseline, GPU Operator and Dynamo health, DRA access, gang scheduling, autoscaling, metrics and gateway readiness; performance validation runs separately because it consumes GPUs. The pinned versions were tested together only on the listed platforms. Source

4. TensorRT Edge-LLM 0.11.1 Serves Nemotron 3.5 Lightning on DGX Spark, DSpark Speculative Decoding Lifts Throughput Up to 73%

NVIDIA showed how TensorRT Edge-LLM 0.11.1, its lightweight C++ inference runtime for embedded targets, runs on a DGX Spark 64GB as the development box for Jetson Thor, DRIVE AGX Thor and IGX Thor deployments. Installing tensorrt-edgellm[server]==0.11.1 (TensorRT 10.16.1.11 or newer in the 10.x series, CUDA 13, Python 3.10 to 3.12) provides trtedge chat, which downloads a model and builds engines, and trtedge serve, which exposes an OpenAI-compatible /v1/chat/completions endpoint with streaming plus a /health/ready check. With Nemotron 3.5 Lightning 30B-A3B in NVFP4 (about a 21GB download) and thinking enabled, DSpark speculative decoding with three draft tokens raised generation from 73.3 to 108.5 tokens per second on SpeedBench 1K, 72.8 to 105.3 on SpeedBench 2K and 73.4 to 126.9 on HumanEval prompts, at batch size one with a 128-token output cap on DGX OS 7.6.0. Median time to first token ranged from about 98 ms to 332 ms and peak system memory was 27.9GB including the OS. The speculative path needs a separate DSpark draft checkpoint, and the model is text-only. Source