Interactive AI inference
Performance models and execution strategies for frontier language models, speculative decoding, long-context inference, and mixture-of-experts architectures.
MODEL × SYSTEMAMD RESEARCH & ADVANCED DEVELOPMENT · AUSTIN, TEXAS
Vignesh
Adhinarayanan
Architecture and systems for large-scale AI inference.
I research GPU architecture, high-bandwidth memory, interconnection networks, data orchestration, and parallel execution models for AI accelerators.
20+peer-reviewed publications
13U.S. patent grants
3international patent grants
2×ISCA publications
RESEARCH AGENDA
Co-design across the full inference stack.
My work connects model structure to the physical costs of computation, data movement, memory capacity, and synchronization.
GPU communication
Low-latency collectives, on-chip and scale-out interconnects, and topology-aware mappings for distributed accelerator systems.
COMPUTE × NETWORKMemory systems
High-bandwidth memory organizations and near-memory techniques for fine-grained, irregular, and bandwidth-constrained workloads.
ARCHITECTURE × CIRCUITSSELECTED WORK
Recent publications
Research spanning AI inference efficiency, 3D-stacked memory, and large-scale computing systems.
2026
The qs Inequality: Quantifying the Double Penalty of Mixture-of-Experts at Inference
Vignesh Adhinarayanan and Nuwan Jayasena
2025
Folded Banks: 3D-Stacked HBM Design for Fine-Grained Random-Access Bandwidth
Vignesh Adhinarayanan, Bradford M. Beckmann, Wantong Li, et al.
2025
Looking Back to Look Forward: 15 Years of the Green500
Vignesh Adhinarayanan and Wu-chun Feng
2023
A Research Retrospective on AMD's Exascale Computing Journey
Gabriel H. Loh, Michael J. Schulte, Mike Ignatowski, Vignesh Adhinarayanan, et al.
CURRENT FOCUS