I build performance models that explain how AI workloads use GPU hardware. At NVIDIA, I use those insights to evaluate architecture tradeoffs and guide engineering and product decisions.
My work spans LLM prefill and decode, video generation, and agent-assisted performance analysis. I connect attention, precision, and parallelism choices to system performance, taking studies from model assumptions and simulation through validation and technical recommendations.
I earned my Ph.D. in Computer Science and Engineering at the University of Michigan, advised by Prof. Trevor Mudge. My research connects algorithms and systems through hardware-software co-design, with publications at HPCA, ASPLOS, MICRO, ISCA, VLDB, and PACT.
My experience includes performance modeling for the NVIDIA RTX PRO 6000 Blackwell family and Rubin architecture studies. I focus on how model structure and execution strategy interact with compute, memory, and communication limits.
NVIDIA — Deep Learning Performance Architect
Performance Modeling and Architecture Pathfinding · Santa Clara, CA
NVIDIA — Deep Learning Performance Architect Intern
Santa Clara, CA
Micron Technology — Advanced Hardware Development Intern
Allen, TX
My academic work pairs workload characterization with architecture and system design to improve the efficiency of data-intensive computation. Selected papers below cover recommendation inference, GPU and memory systems, graph processing, and LLM fine-tuning. See Google Scholar for the full publication list.
Palermo co-designs an oblivious RAM protocol and hardware controller to overlap memory operations while preserving correctness and security guarantees. The paper evaluates how protocol and architecture choices reduce the performance cost of protecting memory access patterns.
GRACE clusters an item co-occurrence graph and caches embedding partial sums to reduce recommendation inference memory traffic. It pairs an algorithmic approach to recurring item combinations with a scalable system design.
This work characterizes single-GPU fine-tuning of sparse and dense mixture-of-experts language models. An analytical model estimates throughput and cloud training cost to support infrastructure choices.
Everest accelerates temporal motif mining on GPUs through query-specific code generation, runtime load balancing, and multi-GPU graph partitioning. It connects graph-query structure to efficient parallel execution.
This work introduces locality-adaptive remote memory transfers and a remote-only cache for multi-GPU systems. The architecture targets remote access latency by matching transfer and caching behavior to workload locality.
NDMiner combines near-data execution, compiler support, and memory mapping to accelerate graph pattern mining. The design addresses the irregular memory access and control flow that make this workload difficult to execute efficiently.
Mint introduces an accelerator architecture and programming model for temporal motif mining. It exploits parallelism while handling both structural and temporal constraints in graph queries.
Sparse-TPU co-designs sparse matrix packing and a systolic-array architecture to improve sparse matrix-vector multiplication efficiency. The work examines how data representation and hardware support can make better use of array resources.
Ph.D., Computer Science & Engineering
University of Michigan, Ann Arbor · Advised by Prof. Trevor Mudge
GPA 4.0/4.0 · Focus: Computer Architecture for AI workloads
B.S.E., Computer Engineering
B.S.E., Electrical & Computer Engineering
