Project Grant 2544462

The National Science Foundation Division of Information and Intelligent Systems awarded $384,296 to the University of California, San Diego on August 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop a disaggregated, heterogeneous context mixing plane for scalable generative AI systems.

The research addresses inefficiencies in current transformer-based AI architectures by separating memory-bound context mixing computation from compute-bound dense computation, allowing each to scale, evolve, and fail independently. The project organizes work into three thrusts: designing and validating a fault-tolerant attention-as-a-service runtime for long-context language model training, multi-model serving, and reinforcement learning; co-designing context mixing algorithms for emerging workloads including video diffusion and multimodal generation; and implementing the system on heterogeneous accelerators through topology-aware communication libraries, parallelism-aware scheduling, and kernel compilation. The research aims to reduce the cost and energy of training and serving long-context AI models by an estimated order of magnitude, lowering barriers for researchers at smaller institutions and academic agencies while broadening the accelerator market beyond single vendors.

Performance occurs in La Jolla, California, with the award period extending from August 1, 2026, through July 31, 2031.