Penn Launches Betty NVIDIA Supercomputer for AI Research
The University of Pennsylvania (Penn) is expanding its artificial intelligence and high-performance computing capabilities with a new off-campus supercomputer named Betty.
Developed with NVIDIA and operated from a data center in Collegeville, Pennsylvania, roughly 30 miles from Penn’s Philadelphia campus, Betty provides researchers with centralized access to high-performance GPU and CPU resources for increasingly demanding AI and scientific workloads.
The system represents more than a hardware upgrade.
Penn is using Betty as part of a broader shift toward shared AI and high-performance computing infrastructure, allowing researchers from multiple schools and disciplines to access computational resources that would be difficult and expensive for individual laboratories to operate independently.
The initiative also reflects a larger trend across academic research: as AI models and scientific datasets become more computationally intensive, centralized GPU infrastructure is becoming an increasingly important part of the research environment.
🖥️ Betty Brings AI-Scale Computing to Penn Research #
Betty is designed to handle computational workloads that exceed the capabilities of conventional laboratory servers and workstations.
Its NVIDIA-based infrastructure combines high-performance GPUs with CPUs to support workloads involving large datasets, model training, inference, simulation, and other computationally intensive research tasks.
The system can provide researchers with significantly more compute capacity than a typical lab-owned server while allowing the underlying infrastructure to be shared across multiple research groups.
Kenneth Chaney, Associate Director of AI and Technology at Penn’s Advanced Research Computing Center (PARCC), emphasized that Betty was developed as a collaborative resource spanning multiple areas of the university, including the School of Engineering and Applied Science, the Perelman School of Medicine, and the School of Arts and Sciences.
The underlying motivation is straightforward: modern AI research increasingly requires compute resources that are too expensive or too large for an individual laboratory to maintain.
“The scale of modern AI research has grown beyond what any single school can support.”
A shared system allows researchers to access that capacity without each laboratory having to build its own dedicated infrastructure.
🤝 From Lab-Owned Servers to Shared HPC #
Betty represents a transition away from the traditional model in which individual laboratories purchase and manage their own servers.
That model can work well for specialized workloads, but it also creates an inefficient utilization pattern.
A research lab may purchase a powerful GPU server for a particular project, yet the machine may remain idle when the researchers are not actively running experiments.
A centralized high-performance computing environment changes that equation.
Instead of distributing expensive hardware across dozens of laboratories, Penn can pool the resources into a shared platform and allocate compute capacity according to demand.
This approach can improve:
- GPU utilization
- Infrastructure efficiency
- Hardware lifecycle management
- Research accessibility
- Data center operations
- Collaboration between research groups
Marylyn Ritchie, Vice Dean of AI and Computing at the Perelman School of Medicine, described this as an important advantage of centralized computing.
“Most lab computers aren’t running 24/7; they often sit idle.”
With Betty, the same hardware can support different research teams at different times, allowing the university to extract more value from expensive GPU infrastructure.
⚡ Why Betty Is Located Off Campus #
Betty is housed in a data center in Collegeville rather than directly on Penn’s Philadelphia campus.
The primary reason is infrastructure.
A large AI supercomputer requires considerably more power and cooling than a conventional laboratory server cluster. Betty requires approximately 1 megawatt of electrical power, along with specialized thermal and facility infrastructure.
Providing that capacity in an urban campus environment would be significantly more complicated.
An external data center can provide:
- High-capacity electrical infrastructure
- Specialized cooling
- Higher rack density
- Dedicated facility operations
- Scalable power delivery
- Physical infrastructure designed for HPC systems
Moving the hardware off campus therefore allows Penn to deploy a much larger computing platform without requiring the university to construct an equivalent data center inside Philadelphia.
The important point is that the system remains logically connected to the university.
Researchers can access Betty through Penn’s network, meaning the physical distance between the compute cluster and campus does not prevent researchers from using it.
“As long as you can connect to Penn’s network, you can access Betty.”
This is increasingly common for modern research computing. Physical proximity to the hardware is less important when high-speed networking and remote-access infrastructure are available.
🧠 AI Computing Is Becoming a Cross-Disciplinary Resource #
One of the most important aspects of Betty is its intended user base.
The system is not designed exclusively for computer science or AI researchers.
Penn expects GPU computing to become increasingly important across disciplines.
Ritchie argues that while GPU acceleration is currently critical for certain research areas, AI and accelerated computing will eventually become much more broadly applicable.
Potential research areas include:
- Computational biology
- Drug discovery
- Medical imaging
- Protein and biomolecular research
- Natural language processing
- Computer vision
- Robotics
- Scientific simulation
- Data-intensive research
- Large-scale machine learning
This shift is important because GPUs are no longer simply graphics processors.
Modern accelerators have become general-purpose computational engines for workloads involving massive parallelism, matrix operations, and machine learning.
As a result, the same underlying infrastructure can potentially support researchers working on very different scientific problems.
🧬 Early Research Projects #
Betty entered pilot operation with 10 research laboratories and 47 researchers already using the system.
Early projects include several examples of how centralized AI infrastructure can support different research domains.
Project Eureka #
Led by Dinesh Jayaraman, Assistant Professor of Computer and Information Science, Project Eureka represents one of the AI-focused research efforts using Betty.
Large-scale experimentation in machine learning often requires repeated training runs, extensive datasets, and significant accelerator time.
A shared GPU cluster can make these experiments substantially easier to execute at scale.
Large-Scale Models for Biomolecular Methods #
Another project focuses on large-scale models for biomolecular research and is led by Pranam Chatterjee, Assistant Professor of Bioengineering and Computer & Information Science.
This type of workload illustrates the growing convergence between AI and biological research.
Modern biomolecular models can involve large datasets and computationally intensive training and inference, making GPU infrastructure increasingly useful for researchers outside traditional computer science departments.
NSF AIRFoundry #
Betty is also being used by NSF AIRFoundry, led by Zack Ives, Department Chair and Adani President’s Distinguished Professor of Computer and Information Science.
The project further demonstrates how shared computing resources can support research programs that operate across different computational domains.
🏛️ Honoring Frances “Betty” Holberton #
The name Betty is also a reference to Penn’s computing history.
The system is named after Frances “Betty” Holberton, one of the six original programmers of ENIAC, the electronic digital computer developed at the University of Pennsylvania during the 1940s.
Holberton played an important role in the early development of programming techniques for ENIAC.
Naming the new supercomputer after her creates a direct connection between Penn’s early contributions to electronic computing and its current investment in AI infrastructure.
The historical connection is particularly fitting.
ENIAC represented a radical increase in computational capability for its era.
Betty represents a similar transition in a very different technological context: the movement from conventional research computing toward large-scale GPU-accelerated AI systems.
🔬 Why Shared GPU Infrastructure Matters #
The Betty project highlights a broader architectural change taking place across universities.
Historically, scientific computing often revolved around CPU-based clusters and specialized laboratory servers.
AI has changed the resource profile.
Modern workloads can require large numbers of GPUs, high-bandwidth memory, fast interconnects, substantial power delivery, and sophisticated cooling systems.
Those requirements make centralized infrastructure increasingly attractive.
A shared AI supercomputer can provide researchers with access to resources that would otherwise be difficult to justify at the individual-lab level.
The model also creates a natural path toward better utilization.
Instead of dozens of partially utilized GPU servers distributed across campus, a centralized system can schedule workloads dynamically and allow different groups to share the same computational pool.
This is essentially the same economic principle behind cloud computing, applied to academic research infrastructure.
🌐 AI Research Is Becoming Infrastructure-Driven #
Betty also illustrates a broader trend in AI research: the limiting factor is increasingly infrastructure rather than software alone.
Advanced research depends on a stack that includes:
- AI models
- Training frameworks
- Accelerators
- High-bandwidth memory
- High-speed networking
- Distributed storage
- Power delivery
- Cooling
- Scheduling software
- Data management
- Research expertise
Universities that want to remain competitive in AI therefore need more than talented researchers.
They also need access to increasingly sophisticated computing infrastructure.
For Penn, Betty provides a centralized foundation for that infrastructure while reducing the need for individual research groups to independently acquire and maintain expensive GPU systems.
🚀 Betty Is a Foundation for Penn’s AI Future #
The launch of Betty marks a significant step in Penn’s approach to AI and high-performance computing.
Rather than treating accelerated computing as an isolated resource for a small number of specialized laboratories, the university is positioning GPU infrastructure as a shared research platform.
Its off-campus location provides the electrical and cooling capacity required for a system operating at approximately 1 MW, while network connectivity keeps the resource accessible to researchers across Penn.
With 10 laboratories and 47 researchers already participating during the pilot phase, Betty is effectively serving as an early test of this shared infrastructure model.
The larger implication extends beyond one university.
As AI becomes embedded in biology, medicine, engineering, computer science, and the broader scientific research ecosystem, access to scalable GPU computing will increasingly become part of the basic infrastructure of modern research.
From ENIAC to Betty, Penn’s computing story continues to evolve—from building some of the world’s earliest electronic computers to deploying specialized infrastructure for the next generation of AI-driven science.
