MRI Consortium: Track 1: Development of SPECTRUM: An Evolutionary Rapid-Prototyping Testbed for Low Latency Multi-Domain Computing

About This Grant

As artificial intelligence (AI) expands into fields such as healthcare, robotics, and energy distribution, there is growing demand for faster, cost-effective ways to prototype AI hardware. Graphics processing units (GPUs) offer high throughput but often struggle to meet the low latency needed for real-time decisions. Field-programmable gate arrays (FPGAs) provide lower latency but can substantially sacrifice throughput. The key challenge is mapping AI algorithms to the right accelerator. The SPECTRUM project establishes a testbed that combines GPUs and FPGAs, both equipped with embedded tensor processing cores, to help reduce latency. Using the CHARM tool, developers can use the SPECTRUM testbed to quickly test hardware setups for real-time AI systems like autonomous vehicles. SPECTRUM advances the state of AI system prototyping, particularly for low latency designs, through the development of its reconfigurable testbed composed of FPGAs (e.g., AMD Versal ACAPs) and GPUs (e.g, Nvidia Hopper/Blackwell GPUs), which contain embedded tensor cores. The CHARM flow automatically partitions the AI computation between the traditional accelerator hardware and the embedded tensor cores. The work is organized into three thrusts: (1) acquisition and integration of hybrid accelerators supporting both edge and data-center configurations; (2) extension of the CHARM HW/SW framework to enable automated, domain-specific accelerator synthesis across heterogeneous platforms; and (3) creation of a user-facing SPECTRUM application interface for deployment and evaluation. The infrastructure enables scalable, low-latency (<10ms), end-to-end AI accelerator design, significantly lowering the barrier to entry for domain experts in real-time, safety-critical applications. SPECTRUM has so far developed an interested user base from research efforts across 26 groups at 16 institutions, delivering critical national infrastructure for real-time AI system design. Access will be provided directly from the lead site at Syracuse University and integration is planned into the FABRIC network. A key outreach strategy includes hands-on workshops and tutorials at major conferences in supercomputing, computer systems, design automation, and FPGAs to train a broad user base. Online courses will support workforce development across skill levels. The testbed will be integrated into undergraduate and graduate curricula at Syracuse, Brown University, and University of Pittsburgh, and will support interdisciplinary collaborations and K–12 outreach initiatives focused on AI hardware and real-time computing. To support long-term community engagement and reproducibility, the project team will develop and maintain a public-facing website at https://spectrum-ai.org, which will serve as the central hub for documentation, datasets, tutorials, benchmark results, and application deployment workflows. All software tools, including the CHARM framework, will be hosted on GitHub and linked from the main site. The project team will ensure that the website and repositories remain accessible and maintained for at least five years beyond the operational lifetime of the physical testbed, supporting continued use by researchers, educators, and students in real-time AI system design. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.