University of Pennsylvania Accelerates Research Across Disciplines With NVIDIA DGX SuperPOD

The rise of large-scale AI has fundamentally changed what it means to do computational science. Fields such as biomedical science, robotics, linguistics, and materials science that require specialized computing resources now generate training workloads that push against the limits of what any single lab can practically support. For Penn, a university with twelve included schools pursuing federally funded research simultaneously, the question was not whether to invest in AI infrastructure but how to build something shared, accessible, and powerful enough to serve every researcher on campus.

Public cloud could fill short-term compute gaps, but the economics proved difficult at academic scale. Running repeated experiments across dozens of model variants, random seeds, and large datasets is the kind of rigorous work that fosters innovation and shared published results, as well as grant opportunities for further exploration. But it required a level of compute that was either unavailable locally or costly enough to make iteration impractical. The most ambitious experiments were hardest to pursue precisely when they mattered most: before results existed to justify additional economic support for advancement.

What Penn needed was a platform that could meet every researcher where they were—scalable enough for frontier model training, affordable enough for broad access, and integrated enough that getting started took hours, not months.

The impact on research scope and velocity has been immediate. In Penn's robotics program, researchers running reinforcement learning experiments for navigating unknown environments can now fully parallelize training trials on Betty. The same team trained predicate-informed vision-language-action (VLA) models at 3B+ parameters previously out of reach due to compute requirements. The results achieved a 12x speedup in model training compared to their prior setup, with preliminary VLA results that outperformed the state of the art. "On a desktop GPU, each reinforcement-learning training run takes three to four hours, limiting us to six or seven runs a day. With just a fleet of eight GPUs on our DGX SuperPOD, we can complete 100 runs with different configurations per day—more than a 14x increase in daily experimental throughput," said Vijay Kumar, Professor and Nemirovsky Family Dean, Penn Engineering.

Betty’s impact extends well beyond any single lab. A collaboration between the Mack Institute and Penn Center for Innovation used Betty to analyze semantic correlations across more than 40,000 documents, reducing an estimated 1.5 years of compute time to just three days. CAMEL-ECG, a model designed to reduce alarm fatigue in clinical ECG diagnostics, was trained on Betty by a team spanning Penn Engineering and the Perelman School of Medicine: a collaboration enabled directly by shared access to the platform.

Biomedical science is a long-standing strength at Penn, with researchers who have pioneered Nobel Prize-winning mRNA vaccine technology. Betty now supports a generative model for mRNA and protein constructs that has demonstrated over 400-fold higher expression. Penn faculty have also crawled the full PubMed repository and extracted a knowledge graph for training predictive models across biomolecule types, with an agentic co-scientist system built on top of that data to support RNA and lipid nanoparticle therapeutics development.

Today, Betty serves approximately 950 registered researchers across 200 research groups, delivering 120,000 GPU hours and 30,000 jobs per month at a subsidized cost of roughly $1 per GPU hour. Any Penn researcher can request access, and new users can move from sign-up to productive work on Betty within hours rather than weeks. “Betty gives Penn researchers integrated AI capacity previously unavailable on campus. At 10x+ lower cost than commercial cloud alternatives, the DGX SuperPOD turns large-scale model training and AI-for-science projects from nonstarters into research teams that can build, test, and refine,” Vijay added.

Penn is already expanding Betty's model portfolio. The platform now serves frontier-scale open models, including GLM-5.2-NVFP4—an NVIDIA-optimized deployment for NVIDIA Blackwell GPUs supporting context lengths up to 500,000 tokens with multi-node Kimi K3 deployments underway. Betty’s capabilities support not only research but also teaching. Penn's new "Software Development with AI Coding Agents" course launches this fall with 75 students using a Betty-hosted model endpoint for coding assignments and grading, within Penn's data governance boundaries, with enrollment expected to grow to 250 students in the spring.

In biomedical imaging, Penn researchers are building multimodal datasets spanning from organ scale to single-cell resolution, creating the foundation for models that could eventually predict organ function or support regenerative medicine. The clinical impact of Betty's work continues to emerge, but the platform already drives the computational workloads that select peptide, protein, and mRNA candidates for experimental testing and potential translation into treatment.

Penn's investment in Betty demonstrates what becomes possible when a research university treats compute not as a departmental resource but as a shared foundation for innovation—open to every researcher, affordable at academic scale, and powerful enough that the science, not the hardware, sets the limit.