NVIDIA AI Updates: October 24, 2026

1. TensorRT 11.1 Global Performance Tuner Picks the Fastest Engine That Meets an Accuracy Tolerance

NVIDIA. NVIDIA published a two-part technical series on the Global Performance Tuner, an automatic performance-tuning feature released in TensorRT 11.1. Instead of building an engine through a single fixed scheduling route (graph transformations, fusion, kernel selection, memory placement and quantization), the tuner generates many candidate engines through customizable build routes. It then checks each candidate against end-to-end accuracy, measures end-to-end inference performance, and selects the fastest engine that stays within the accuracy tolerance. Part 2 covers the algorithm, architecture and best practices. Source Source

2. European Banks Deploy Open Models on NVIDIA Infrastructure, From Revolut’s PRAGMA to Santander’s Kumo Pilot

NVIDIA. At GTC Berlin, NVIDIA highlighted European financial institutions running open-weight and in-house models on its platform. Revolut is scaling its PRAGMA foundation model on Blackwell Ultra GPUs with cuDF and Dynamo-Triton for an initial 80 million customers, and Lloyds Banking Group built the ATUM transaction foundation model on RTX PRO 6000 Blackwell GPUs. Neobank bunq said its Finn assistant now runs on open-weight models on Nebius at 60% of the cost per answer while matching leading closed models, Modulr is building payment models on Nemotron-3-Ultra-550B, and Santander is an early adopter of the NVIDIA Kumo open models for structured financial data. LSEG also said the Polars GPU engine with cuDF analyzes historical tick data up to 10x faster. Source

3. Shopify Post-Trains Nemotron 3.5 Lightning, Cutting Sidekick Subagent Serving Costs by 90%

NVIDIA. NVIDIA published a roundup of European retail deployments of its open models and physical AI stack. Shopify post-trained Nemotron 3.5 Lightning with Megatron Bridge and NeMo RL for the GraphQL subagent in its Sidekick merchant assistant, and says the model beats a leading frontier model on its internal benchmark while cutting serving costs by 90% and responding 7x faster. NVIDIA CCluster vLLM optimizations delivered 4.5x faster performance and up to 12x lower inference cost across five more Shopify use cases. Zooplus, working with Infosys, reached 86% search accuracy in a prototype using Nemotron Retriever and the retail catalog enrichment blueprint. Source

4. KvikIO Guide Shows Python Reads Straight Into GPU Memory via GPUDirect Storage

NVIDIA. NVIDIA’s technical blog walked through using KvikIO, a CUDA-X Python and C++ library layered on the cuFile API, to read files directly into GPU memory with GPUDirect Storage. The CUDA 13 build installs with pip install kvikio-cu13, and reads go through kvikio.CuFile(path, "r").read(gpu_buffer) into a CuPy array. The post documents tuning knobs such as compat_mode, gds_threshold, task_size and thread count, plus CUFILE_STATS_LEVEL=3 for checking whether I/O used the accelerated nvidia-fs path or the CPU-staged POSIX fallback. KvikIO is already used inside cuDF, cuCIM and GPU-accelerated Presto. Source

5. Agent-Driven Workflow Trains the Microduck Robot in Isaac Lab and Newton

NVIDIA. NVIDIA described an end-to-end workflow where AI agents plan a robotics project, generate 3D assets (the Astra agent uses FreeCAD and Blender to export OpenUSD files) and help train the small Microduck robot to walk an obstacle course. The pipeline combines Isaac Sim for the course and sensor simulation, the open-source Newton physics engine, and Isaac Lab (with v3.0.0 early-access docs referenced) for reinforcement learning, with Better Actuator Models and domain randomization to narrow the sim-to-real gap. An adaptation run used 4,096 parallel environments on a single RTX PRO 6000 Blackwell Workstation Edition GPU, and the resulting policy was shown walking on a physical Microduck. Source