Vu Le

PhD Student · UMass Amherst · Berkeley Lab

I’m a third-year Computer Science PhD student at the University of Massachusetts Amherst (UMass), working with Prof. VP Nguyen and Prof. Deepak Ganesan. I’m also a research affiliate at Berkeley Lab, hosted by Dr. Yilun Xu, where I worked on computer architecture and heterogeneous computing for sub-microsecond qubit readout.

My primary research interest is in computer architecture, heterogeneous AI computing (GPU, FPGA, NPU), and hardware-aware performance optimization. I study how computation, data movement, dependencies, and execution overhead shape performance, and build analytical hardware models to guide hardware-aware optimizations, including kernel optimization, tiling, dataflow, spatial mapping, operator fusion, and scheduling. My goal is to minimize end-to-end latency under hardware resource constraints for emerging AI-related workloads.

I've evaluated these ideas on spatial AI accelerators and FPGAs (extending to GPUs) for real-time quantum bit readout, transformers, and vision-language-action models. I also contribute high-performance kernels and compiler improvements to AMD’s open-source MLIR-AIR compiler. My earlier work includes edge GPU and FPGA-based AI acceleration.

I am seeking research or industry internships in the US for Summer/Fall 2027 in ML performance engineering and hardware acceleration, ML compilers and kernels, and hardware–software co-design, particularly low-latency inference for physical AI across NPUs, FPGAs, and GPUs. Feel free to reach out!

Get in touch

Personal: vule20.cs AT gmail [DOT] com
UMass: vdle AT umass [DOT] edu

CV (PDF)

Vu Le profile photo

Lederle Graduate Research Center
740 North Pleasant Street, Amherst, MA

Active Research

Analytical modeling scheduler mapping AI workloads onto a spatial NPU

Hardware-Aware Acceleration of VLAs on the Spatial NPUs (abstraction)

Developing hardware-aware performance models and optimization methods for mapping matrix multiplication, attention, and transformer workloads onto spatial AI accelerators. The work focuses on analytical cost modeling, tiling, dataflow selection, spatial mapping, and scheduling under hardware resource and latency constraints. Contributions to AMD's open-source MLIR-AIR compiler include optimized GEMM and FlashAttention kernels, hardware-aware tiling, 2D attention mapping, A-stationary GEMM lowering, and compiler improvements. These compiler improvements significantly accelerate LLMs and SmolVLA; as tracked in PR #2085, they deliver 53.0% lower vision-stage latency and 43.0% lower end-to-end latency compared to AMD's original MLIR implementation, while outperforming both Ryzen AI Software and FastFlowLM.

Analytical modeling scheduler mapping AI workloads across an NPU and a GPU

Hardware-Aware Scheduling of Emerging AI Workloads in a Heterogeneous System

Defining

Extending the hardware-grounded latency model and scheduler beyond a single accelerator to a heterogeneous system. The goal is to partition an AI workload across the two devices or more devices in an SoC, matching each stage to the hardware whose compute and data-movement profile suits it, while keeping end-to-end latency within a real-time budget.

Selected Research

For a full list of publications, please visit my Scholar

Heterogeneous computing pipeline for quantum state classification 2026

Real-Time Recurrent Qubit-State Discrimination on a Heterogeneous AI-Engine–FPGA Accelerator

Vu Le, Yilun Xu, Neel Vora, Gang Huang, Deepak Ganesan, Phuc Nguyen

arXiv 2026 · preprint

The Problem. Conventional I/Q integration discards the temporal information from qubits that relax mid-readout, while full-resolution feed-forward networks (FCNs) are too expensive to deploy under tight real-time constraints. A standard recurrent network (LSTM) captures the dynamics, but its hidden state does not fit in the register capacity of a single AI Engine (AIE) tile and suffers memory-spill latency.

The solution:Heterogeneous Partitioning. We split the LSTM at its dependency boundary. Independent input projections run on the AIE array, using just 4 of 400 tiles. Loop-carried hidden-state updates run on a custom micro-architecture (running at 500MHz) in the tightly coupled Programmable Logic (PL). The accelerator keeps pace with a 64 ns input cadence and, measured end-to-end on a physical VCK190, finishes in 860 ns within the 1,000 ns budget, versus 1,064 ns for a pure-AIE mapping.

Performance. On qubits affected by mid-readout relaxation, accuracy matches a full-resolution FCN (96.59% vs. 96.37%), and a compact variant reaches the highest mean of 96.78% across eight tested qubits. It uses 403× fewer parameters and ~103× fewer DSPs than the FCN, and cuts relative classification error by 68.2% over GMM, 43.6% over QubiCML, and 16.82% over MF-NN. The AIE–PL split reduces exposed tail latency by 60.3% versus pure AIE and DSP58 usage by 31.1× versus an all-PL design.

Detection and Tracking of Drone Swarms using LiDAR 2025

Detection and Tracking of Drone Swarms using LiDAR

Tasnim Azad Abir, Vu Le, Endrowednes Kuantama, Pranjol Sen Gupta, Austin Copley, Judith Dawes, Mohammad Islam, Richard Han, Phuc Nguyen

ACM MobiSys 2025 · A* conference

LiSWARM is a low-cost LiDAR system for accurate 3D tracking and recognition of drones in large swarms. Using point cloud processing, clustering, and neural networks, it achieves up to 98% accuracy and scales to 15,000 drones—enabling applications in airspace security, drone shows, and sensitive area monitoring.

MagicStream immersive telepresence 2024

MagicStream: Bandwidth-conserving Immersive Telepresence via Semantic Communication

Ruizhi Cheng, Nan Wu, Vu Le, Eugene Chai, Matteo Varvello, Bo Han

ACM SenSys 2024 · A* conference

MagicStream is a first-of-its-kind semantic-driven immersive telepresence system that extracts and delivers compact semantic details of users' captured 3D representations, instead of the traditional bit-by-bit communication of raw content.

Fast and Interpretable Face Identification using Vision Transformers 2024

Fast and Interpretable Face Identification for Out-Of-Distribution Data Using Vision Transformers

Hai Phan, Cindy Le, Vu Le, Yihui He, Anh Totti Nguyen

CVF/WACV 2024 · A conference

Uses vision transformers for face identification on out-of-distribution data, running twice as fast as the state-of-the-art DeepFace-EMD model while achieving comparable performance.

News

10/2026

Released our preprint on real-time quantum bit readout: Real-Time Recurrent Qubit-State Discrimination on a Heterogeneous AI-Engine–FPGA Accelerator.

09/2026

My high performance AI kernels are merged into AMD's MLIR-AIR compiler, boosting performance for VLAs and LLMs inference on AMD XDNA2 NPU.

04/2026

Presented research on real-time qubit readout to LBNL and AMD in Berkeley, CA.

06/2025

Presented “Opportunities in Computer Systems Research for Quantum Computing” at ACM QSys 2025 (in conjunction with ACM MobiSys 2025).

03/2025

One paper accepted at ACM MobiSys 2025.

01/2025

I received the James Kurose Scholarship in Computer Science.

12/2024

I officially became a research affiliate with Berkeley Lab.

Recent Collaborators

I’m fortunate to have worked with Hai Phan (Woven by Toyota), Ethan He (xAI), Bo Han (George Mason University), Ruizhi Cheng (Amazon), Tasnim Azad Abir (UTA), Richard Han (Macquarie University), Yilun Xu (Berkeley Lab), Amir Radmehr (UMass Amherst)

Research Supports

Miscellanea

Outside of work, I'm an avid adventurer — I love road trips and have driven across the US to explore national parks and trails firsthand. Most of my hikes and adventures (Grand Canyon, Zion, Death Valley, Horseshoe Bend, Joshua Tree, the White Mountains in New Hampshire) were made possible by hitting the road. I also really enjoy lifting at the gym, jogging, brewing coffee, planting flowers, and skiing. I shoot with a Sony A6400 and a collection of Sigma and Sony lenses. Check out my photo gallery.