Efficient Generative & Agentic AI
Efficient inference for diffusion language models, KV-cache optimization, and system support for emerging agentic workloads.
Polestar · Polestar-Cache
Mingyu Lee
Ph.D. Student in Electrical and Computer Engineering
Georgia Institute of Technology · Synergy LabKlaus Advanced Computing Building · Room 3306 · Atlanta, GA
I am a Ph.D. student at Georgia Tech, advised by Prof. Tushar Krishna. My research focuses on efficient AI systems through cross-layer hardware-software co-design.
I work across algorithms, systems, and hardware to make emerging models practical under real deployment constraints. My current interests include efficient generative and agentic AI, model compression, emerging data formats, and accelerator architecture.
Research Interests
My research asks how model structure, numerical representation, runtime systems, and hardware can be designed together rather than optimized in isolation.
Model Compression
Data-free quantization, structured activation pruning, sparsity, and emerging low-precision data formats.
OuroMamba · RECAP
Hardware-Software Co-Design
GPU kernels, FPGA/ASIC accelerators, systolic arrays, RISC-V extensions, and memory-efficient architectures.
CUTLASS · FPGA · 65 nm ASIC
News
Started my Ph.D. in ECE at Georgia Tech 😊!
Released Polestar, our work on efficient diffusion LLM inference.
Presented MiniSA at ISPASS 2026 on behalf of Jianming.
Polestar-Cache appeared at the LIT Workshop at ICLR 2026.
Presented OuroMamba at ICCV 2025 in Honolulu.
Presented RECAP at MLArchSys, co-located with ISCA 2025.
Selected Publications
Explore by research area · select multiple areas to find work at their intersection. · * Equal contribution
Showing all 4 publications
OuroMamba: A Data-Free Quantization Framework for Vision Mamba Models
Akshat Ramachandran*, Mingyu Lee*, Huan Xu, Souvik Kundu, and Tushar Krishna
ICCVIEEE/CVF International Conference on Computer Vision2025Conference
Data-free calibration and dynamic mixed-precision quantization for Vision Mamba models, with up to 2.36× GPU speedup.
RECAP: Training-Free Compensation for Coarse Activation Channel Pruning in Compressed LLMs
Mingyu Lee*, Akshat Ramachandran*, and Tushar Krishna
MLArchSys @ ISCAWorkshop on ML for Computer Architecture and Systems2025Workshop
A lightweight compensation method for hardware-friendly channel pruning, improving BoolQ accuracy by 34% at 70% sparsity on LLaMA3-8B.
Polestar-Cache: Reconciling Parallel Decoding and Accuracy in Diffusion LLMs via Token Drift-Aware KV Cache Recalibration
Mingyu Lee*, Akshat Ramachandran*, Souvik Kundu, and Tushar Krishna
LIT @ ICLRLearning on Information and Tokens Workshop2026Workshop
A training-free, drift-aware KV-cache framework that selectively refreshes stale token representations during parallel diffusion decoding.
Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs
Mingyu Lee*, Akshat Ramachandran*, Souvik Kundu, and Tushar Krishna
arXivPreprint2026Preprint
A unified drift signal for selective KV-cache refresh and reliable token commitment, improving both throughput and generation accuracy.
No publications span all selected research areas.
Selected Projects
Implementation and hardware work, organized around my specific role and technical contribution.
65 nm Systolic GEMM Accelerator
Role
Designer · RTL to tapeout
TSMC 65 nm · SystemVerilog · Synopsys
My contribution
Designed and taped out a 5.56 TOPS/mW weight-stationary GEMM accelerator, including memory, external, and LFSR-based built-in self-test modes.
CXL-based LLM Accelerator
Role
FPGA Engineering Intern · HyperAccel
CXL · RISC-V · FPGA · FP8
My contribution
Extended the RISC-V ISA and designed adaptive processing elements for Qwen1.5-MoE, reducing model memory footprint by 3× through FP8 quantization.
Sparse ViT Accelerator
Role
Undergraduate Researcher · CAST Lab
FPGA · Sparse data formats
My contribution
Developed a compressed sparse diagonal format and its accelerator architecture, achieving 2× speedup and compression over conventional formats.
Talks & Presentations
ISPASS 2026 · Seoul
Presented MiniSA at the main conference on behalf of Jianming.
ICCV 2025 · Honolulu
Presented OuroMamba at the main conference poster session.
MLArchSys @ ISCA 2025 · Tokyo
Oral and poster presentation of RECAP.
Education
2026 — Present
Georgia Institute of Technology
Ph.D. in Electrical and Computer Engineering
Advisor: Tushar Krishna
2019 — 2026
Korea Advanced Institute of Science and Technology
B.S. in Electrical Engineering
Experience
Synergy Lab · Graduate Research Assistant
2025 — PresentGeorgia Tech
Synergy Lab · Undergraduate Researcher
2025 — 2026Georgia Tech
FPGA Engineering Intern
2024HyperAccel
Undergraduate Researcher
2024CAST Lab, KAIST
Academic Service
2025 · Reviewer
MLArchSys @ ISCA
Workshop on ML for Computer Architecture and Systems.
2025 · Volunteer
CRNCH Summit
Center for Research into Novel Compute Hierarchies.
Honors
2026Georgia Tech ECE Fellowship
2025ICCV Broadening Participation Award
2025uArch Workshop / ISCA Grant
2025HPCA Student Travel Grant
2024Korea–U.S. Student Exchange Scholarship
Beyond Research
I recently started drawing manga, a hobby I hope to keep developing throughout my Ph.D. It has been a welcome way to slow down, observe details, and make something by hand.
I also love traveling. I enjoy making memories with friends, but I also like traveling alone with a camera. Recent trips include Tokyo, the Grand Canyon, and Yosemite in 2025, and Hiroshima in 2026.