Daily News · 2 min read

NVIDIA AI Updates: August 27, 2026

nvidia nvlink-fusion nvhbm cuda python inference robotics

NVIDIA. NVIDIA introduced NVHBM, a custom high-bandwidth memory technology that integrates the memory controller into the HBM base die instead of the compute chip. NVIDIA said the design delivers up to 30% greater memory bandwidth and 15% lower HBM power consumption versus standard HBM4E, while freeing up to 25% more area on the compute die for processing. Amazon’s Annapurna Labs is the first partner to collaborate on NVHBM alongside NVLink Fusion, integrating it with next-generation Trainium4 chips in common rack-scale architectures. Source

2. CUDA Python 1.0 Establishes Stable APIs on a Shared Foundation

NVIDIA. NVIDIA released CUDA Python 1.0, marking Python as an officially supported path to the CUDA platform with semantic versioning guarantees across its components. The release bundles cuda.core 1.0.0 for a Pythonic runtime interface, cuda.compute 1.0.0 for parallel algorithms such as sort, scan, and reduce, and cuda.bindings 13.3.0 for 1:1 coverage of CUDA C APIs, plus supporting tools like cuda-pathfinder and nvmath-python 1.0. The shared foundation lets Numba kernels, cuda.compute calls, and external libraries operate on the same GPU buffers without data copying. Source

3. Qwen3.8-Flash-Next Runs on GB300 NVL72 With Day 0 Support

NVIDIA. NVIDIA detailed running Alibaba’s Qwen3.8-Flash-Next, a 125B-parameter mixture-of-experts model with 6B parameters activated per token and a native 262,144-token context window, on its GB300 NVL72 system. The company reported over 16,000 tokens per second per GPU and over 200 tokens per second per user, enabling interactive agentic coding at scale. NVIDIA said the model has Day 0 functional support through SGLang, vLLM, and TensorRT LLM, with fine-tuning available via NeMo AutoModel. Source

4. NVIDIA Details Agent-Driven Cross-Embodiment Robot Navigation Training

NVIDIA. NVIDIA published a workflow for training navigation policies that transfer across different robot embodiments using its COMPASS framework, which adapts the pretrained X-Mobility foundation policy into residual specialists through reinforcement learning. The pipeline runs on Isaac Lab 3.0 and Isaac Sim 6.0, uses Omniverse NuRec to convert real-world captures into simulation-ready reconstructions, and applies cuVSLAM for camera-based visual odometry. An AI coding agent orchestrates environment validation, training, evaluation, and checkpoint promotion, with human approval gates at key stages. Source