Friday, November 20, 2026 Edition No. 244
Search
Daily News Fully automated · 2 min read

NVIDIA AI Updates: November 18, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. NVIDIA Opens SC26 With Nemotron for Science, Agent Showcases and a Quantum-Classical Co-Design Platform

NVIDIA used Ian Buck’s SC26 opening in Chicago to announce Nemotron for Science, an open model and chemistry dataset that researchers can run locally, alongside an expanded Vera CPU platform for high-performance computing. Showcased agents included a KTH agent that fixed 81% of broken GROMACS test cases in one overnight run on a single DGX Spark, and the Nsight AI Porting Agent, which sped up a Weather Research and Forecasting model routine more than 5x using about 150 million tokens. NVIDIA, IBM and Fermilab also announced a quantum-classical co-design platform and a CUDA-Q decoding algorithm called TCBP-JOBS that cuts error-correction research computation from 12 weeks to 45 minutes. Source

2. NVIDIA Details Co-Designing Sparse Attention for Long-Context Inference on GPUs

NVIDIA published a technical post on co-designing model-level sparse attention, where each query attends to only k keys, for long-context inference. The post notes that the indexer still scores all N keys, so in DeepSeek V3.2 the indexer and Top-k take 47% of prefill layer time at 128K tokens. It recommends using one indexer key head with a cross-head reduction (decode runtime rose 7.6x going from 1 to 8 key heads) and fusing Top-k into the indexer, which with reuse of the prior decode step’s selection is 2.1x to 2.4x faster. Source

3. NVIDIA Shows How Slurm HRES Plugins Schedule QPUs as Cluster Resources

NVIDIA described how Slurm 26.11 Hierarchical Resources (HRES) plugins let supercomputing sites schedule a quantum processor as a cluster-wide resource tied to the nodes that can reach it. A short Lua script handles health checks, admits jobs before nodes are allocated and releases the device when jobs end, while site-specific settings live in a separate resources.yaml file. QRMI and QDMI have released open-source plugins on these APIs, and users request a device with a flag such as --resources=qpu:1. Source