SLO-aware GPU Frequency Scaling for Energy Efficient LLM Inference Serving

A new quantization technique called Four Over Six (4/6) has been introduced to enhance the NVFP4 quantization algorithm, which is crucial for large language models (LLMs). This method evaluates two potential scale factors for each block of values, addressing the performance degradation often seen during training and inference due to quantization errors.

The NASA Astrobee has successfully demonstrated the first on-orbit application of reinforcement learning (RL) for autonomous control aboard the International Space Station (ISS). This achievement involved training a deep neural network using NVIDIA's Omniverse physics simulator, allowing the Astrobee to navigate effectively in microgravity environments. The results validate a new training pipeline that bridges the simulation-to-reality gap, showcasing the potential for RL in space robotics.

The introduction of Comba, a novel variant of Bilinear RNNs, leverages closed-loop control theory to enhance recurrent memory management, presenting a scalar-plus-low-rank state transition model. This development builds on recent advancements in sequence modeling, including Gated DeltaNet and RWKV-7, which have improved performance through innovative memory supervision techniques.

An end-to-end pipeline has been developed for deploying reinforcement learning (RL) trained Artificial Neural Networks (ANNs) on Intel's Loihi 2 neuromorphic hardware, converting them into spiking Sigma-Delta Neural Networks (SDNNs). This innovation was tested using an RL policy for controlling the Astrobee free-flying robot, demonstrating low-latency and energy-efficient inference in NVIDIA's Omniverse Isaac Lab simulation environment.

NVIDIA has announced the open-sourcing of its reasoning model for autonomous driving, named AR1, at the NeurIPS 2025 conference. This model is designed to break down scenes step by step, evaluating possible trajectories and uncertainties to enhance decision-making in self-driving vehicles.