NVIDIA AI Updates: July 31, 2026

1. NVIDIA Releases nvmath-python v1.0 for Core Math at Scale

NVIDIA released nvmath-python v1.0, a Python library that provides optimized access to its CUDA-X math libraries across CPU, GPU, and distributed multi-node systems. The library serves as a Pythonic abstraction layer over the CUDA and NVPL math libraries, including cuFFT, cuBLAS, and cuTENSOR, and integrates with NumPy, CuPy, and PyTorch arrays. New capabilities include support for universal sparse tensors, composite kernel fusion, stateful APIs that amortize planning costs through reuse, and integration with JIT compilers such as Numba-CUDA. Source

2. NVIDIA AI Red Team Outlines Four Controls for Securing AI Agents

NVIDIA published guidance from its AI Red Team outlining four architectural controls for deploying AI agents securely at enterprise scale. The recommendations cover restricting each agent to explicitly authorized users, running agent code in sandboxed environments such as Docker or NVIDIA OpenShell, enforcing a default-deny network egress policy with a least-privilege allowlist, and keeping credentials out of agent environments by using dedicated secret managers. The team argues that controls enforced outside the model’s control plane are essential because prompt-based defenses alone remain vulnerable to social engineering and manipulation. Source

3. NVIDIA Shares Exemplar Cloud Lessons on AI Infrastructure Performance

NVIDIA detailed lessons from its Exemplar Cloud validation program, which requires deployments to reach 95 percent of reference architecture performance on identical GPU systems. The post identifies four recurring configuration issues behind performance gaps: virtualization-related SMMU overhead on Grace CPUs, CPU power management and NUMA misplacement, insufficient NCCL queue-pair concurrency on high-bandwidth fabrics, and missing topology files in containerized environments. Case studies show that fixes to kernel, hypervisor, BIOS, and NCCL settings recovered between 9 and 53 percent of lost performance. Source