Yes, developers can extensively customize Vera Rubin's software stack. The platform is designed with an open, modular, and composable software architecture that allows developers to combine and adjust components without needing to rebuild entire systems. This flexibility is a core tenet of the Vera Rubin platform, enabling a wide range of customization options for various AI workloads, from large-scale training to real-time inference and agentic AI.
At the heart of this customizability is NVIDIA's comprehensive AI software ecosystem. The NVIDIA NeMo Framework, for example, offers an end-to-end workflow for building, adapting, aligning, and deploying large AI models at scale. Developers can leverage NeMo for data curation, distributed training, and parameter-efficient customization. Tools like NVIDIA NeMo Run provide the capability to configure, launch, and manage experiments consistently across different environments, including local setups, SLURM clusters, and Kubernetes-based AI factories. Furthermore, NVIDIA AI Workbench serves as a unified toolkit, simplifying the process for developers to create, test, and customize pretrained generative AI models using their own data, and integrating with popular repositories such as Hugging Face and GitHub.
The Vera Rubin platform's software stack also supports higher-level frameworks like PyTorch and JAX, with native NVIDIA acceleration to enhance developer productivity and scalability. The platform includes a collection of software libraries for partners and developers to build upon, such as DSX Max-Q, which helps optimize computing output and token performance per watt within fixed power budgets. Additionally, Red Hat is actively developing a custom version of its flagship operating system, Red Hat Enterprise Linux (RHEL), specifically designed for Vera Rubin, ensuring day-zero support and further enabling software-level customization for the platform's unique hardware components. For managing and processing vast amounts of unstructured data, the NVIDIA AI Data Platform offers a customizable reference design that integrates accelerated computing and AI software, enabling the creation of semantic knowledge layers that AI applications can query in near real-time. This robust and open software environment empowers developers to tailor Vera Rubin to their specific needs, whether they are working on foundational models, specialized AI agents, or complex inference applications.