NVIDIA Academic Grant Program: AI Inference, Agents, and Systems Software

NVIDIA solicits proposals for innovative projects related to generative AI, particularly agentic systems, distributed inference, language models, diffusion models, and multimodal models, with focus on advancing techniques for customizing, enhancing, and operating AI systems.

Overview

• Hardware award: Up to two NVIDIA DGX Spark units (physical hardware shipped to PI)
• Compute resources: Up to 30,000 H100 80 GB GPU hours or equivalent (maximum of eight concurrent GPUs)
• GPU hours expire six months after award; unused hours will be forfeited
• Award amount determined by NVIDIA awards panel
• Required technology integration: Projects leveraging AI models must incorporate NVIDIA models, such as Nemotron™, Cosmos™, or Omniverse™ wherever possible. All others must make extensive use of NVIDIA software and AI distributions, such as NVIDIA NeMo™ microservices, NVIDIA Dynamo, NVIDIA FLARE™, or CUDA-X™ GPU-accelerated libraries

Priority Areas

Techniques for Customizing, Enhancing, and Operating Gen AI:

• Targeted Improvement: Propose and develop novel methods for targeted model quality improvement, including those in the direction of automated synthetic data generation, iterative self-distillation, or self-play
• Explainability: Investigate new approaches to improving model explainability, be it through architectural modifications, model tool use, or automated extrinsic verification
• Reliability: Propose innovations targeting the inherent unreliability of generative AI systems, be it the occurrences of factual hallucinations, failures to follow complex instructions, lack of robustness with respect to input parameter perturbation, or general output non-determinism
• Safety, Security, Privacy: Investigate model alignment techniques and architectural innovations designed to enhance the explainability, robustness, and overall trustworthiness of models and agents. Research areas might include jailbreaking, prompt injection, data exfiltration, memory poisoning, reconstruction attacks, and model inversion
• Inference Efficiency: Propose innovations in inference implementations to ensure efficient generation of high-quality outputs while maintaining model interpretability and safety. This includes model compression methods, such as quantization and pruning, efficient decoding systems, efficient serving and scheduling approaches, novel disaggregation approaches, and quality-enhancing output post-processing techniques. This also includes improvements to fault-tolerance techniques to reduce downtime when a software or hardware error is encountered

Exploring Compound AI Systems:

• Reasoning: Introduce new methodologies for enabling generative AI systems to solve problems on their own accord, either with the help of external tools (e.g., SAT solvers, logic/arithmetic verifiers), aided by reward models, or through self-reflection/self-consistency. The focus is on the research of new techniques, not application of existing techniques to new modalities/fields/problems
• AI Agent Systems and Architectures: Advance the operation of agentic model systems through improvements to efficiency, reliability, or safety. Propose targeted adjustments to existing AI systems to improve efficiency, reliability, or problem-solving performance

Exploring Systems Software for AI:

• Scalable AI Infrastructure: Propose scalable inference systems, including distributed runtimes, scheduling techniques for performance or energy efficiency, resource management frameworks, and fault tolerance techniques. Develop systems for edge AI and/or systems support for AI safety, security, and privacy
• Compilation and Optimization: Introduce optimization techniques or agentic AI flows for parallelization, tiling, fusion, etc.
• Data Management: Develop optimizations for KV-cache management, RAGs, long-term memory, and other states in distributed inference systems
• AI for Systems: Introduce agentic AI techniques to advance beyond the limitations of manual design

Eligibility & Requirements

Eligible Applicants:

• Full-time faculty at accredited academic institutions that award research degrees to Ph.D. students are eligible
• Postdocs and graduate students must work with a full-time faculty member to submit on their behalf

Submission Limits:

• Each person can submit one proposal per quarter, a maximum of four proposals annually
• Each individual applicant is eligible to receive one award per calendar year

Proposal Requirements:

• Proposals must follow the proposal template and should not exceed four pages, not including appendices
• Must incorporate required NVIDIA technologies as specified in Overview section

Selection Process:

• Not all projects that meet eligibility requirements will be selected for an award
• The final award amount will be determined by the NVIDIA awards panel

Recipient Expectations:

• Award recipients should make reasonable efforts to acknowledge the support of NVIDIA Corporation and reference how specific hardware and software contributed to project results
• Recipients will inform NVIDIA of publications, presentations, open-source code and data releases, and speaking engagements that reference the supported project via the NVIDIA academic grant portal
• Failure to report in the portal will influence future award selection
• Must review NVIDIA Academic Grant Program terms and conditions

Timeline

• Submission window: Quarterly cycle 
• Award duration: GPU hours expire six months after award date
• Reporting: Ongoing obligation to report publications and presentations via NVIDIA academic grant portal

How to Submit:

• Follow NVIDIA proposal template (not exceeding four pages, excluding
  appendices)
• Submit proposal through NVIDIA academic grant portal
• Ensure proposal aligns with one or more of the two main research themes specified
• Include plan for incorporating required NVIDIA models and/or software 

Please note: Full RFP is attached in the "More Information" section of this page. Faculty and researchers interested in applying for these opportunities based on technologies developed or disclosed at Vanderbilt must submit their proposals through the CTTC.