Penn Launches Betty AI Platform Using NVIDIA DGX B200 Systems

The University of Pennsylvania accelerated its artificial intelligence research by launching a new computing platform named Betty.

Betty is named after Frances Elizabeth Snyder Holberton, one of six original programmers of the ENIAC computer at Penn and a pioneer in modern software development. The assembly of the full-stack infrastructure required coordinating compute power, networking, data storage, and colocation across a dedicated team of partners.

The NVIDIA DGX-ready colocation partner Flexential delivered approximately one megawatt of colocation capacity in about three months. Inside the system, NVIDIA InfiniBand provides a high-speed fabric to connect the compute nodes. A dedicated dark fibre connects the external facility directly to the Penn campus to offer researchers high-bandwith and low-latency access.

NVIDIA and the DGX SuperPOD-specialized partner AHEAD supplied the AI compute power, networking, and supporting infrastructure. This hardware operates as an integrated platform rather than a collection of separate GPU servers. The platform runs on an NVIDIA DGX SuperPOD equipped with NVIDIA DGX B200 systems.

Data storage served as a critical design consideration for the deployment. Penn installed a VAST Data 15×10 array delivering more than 250 GB/s of read throughput and 60 GB/s of write throughput across multiple petabytes of data, with a consistent data reduction ratio of 1.5 to 1.

Researchers no longer need to align their workflows with data locations. VAST supports highly competing, data-intensive patterns generated by large AI workloads to keep GPU resources continuously supplied.

Penn uses NVIDIA Mission Control to operate and optimize Betty as a unified DGX SuperPOD. Kenneth Chaney, PhD, PARCC associate director of AI and technology at the University of Pennsylvania, explained the operational scope.

“NVIDIA Mission Control provides clusterwide visibility across the entire DGX SuperPOD and combines health and utilization data with GPU performance, memory, Tensor Core, and communication telemetry. With NVIDIA Mission Control, we can spot poorly functioning workloads, help researchers optimize their code, and shorten time-to-solution while returning capacity to the broader Penn community,” Chaney said.

The cluster runs Slurm for traditional high-performance computing workloads and NVIDIA Run:ai for AI-native workloads, including inference services. This setup lets researchers share the same infrastructure while applying the scheduling approach best suited for each specific task.

The deployment also highlighted the value of direct access to NVIDIA artificial intelligence experts. When Penn Evo 2, a genomics foundation model designed for long-context DNA sequence analysis, scaled on the DGX SuperPOD, NVIDIA researchers and engineers worked directly with the university to tune and optimize the model.

This collaboration allowed Penn researchers to transition from a 1-billion-parameter variant to 7-billion and 40-billion models with context windows spanning up to one million nucleotides. These expansions made experiments possible that were previously impractical on campus.