Research

I build CMOS/FPGA systems for probabilistic computing with p-bits (Ising/Boltzmann machines) and distributed architectures for fast sampling, learning, and quantum-inspired optimization. My approach is full-stack — device physics → architectures → algorithms — advancing from single FPGA prototypes to multi-FPGA asynchronous fabrics targeting million-node p-computers.

Research areas

Distributed probabilistic computing

Balanced multi-chip mapping via probabilistic Potts partitioning
Asynchronous, latency-tolerant links that preserve solution quality
Roadmap and measurements toward 100k–1M+ p-bits

Probabilistic Generative AI & ML

Hardware-aware Deep Boltzmann Machines (DBMs) and EBMs
Contrastive-divergence at extreme sweep counts
Neural Quantum States (NQS) for quantum & scientific data

Quantum-inspired optimization & sampling

Sparse fabrics that emulate dense couplings efficiently
Higher-order couplers to improve prefactors and scaling
APT, SQA, and non-local MC for hard instances

Full-stack focus

Device / physical layer

Physics-inspired p-bits (CMOS today; sMTJs next)
Monte-Carlo-faithful device models and noise tuning
Quality tracked by energy / free-energy metrics
Tight device-to-architecture co-design loops

Architecture

Sparse IMs with multiplexed dense interactions
Higher-order interactions (e.g., XORSAT) realized efficiently
Balanced partitioning & async, latency-tolerant links
Chromatic, massively parallel Gibbs updates

Algorithms

SA / PT with non-local moves (ICM) at scale
Simulated Quantum Annealing (SQA)
DBM/EBM training and NQS sampling
Large-sweep, hardware-aware learning loops

Key figures of merit

1500B flips/s (measured) Up to 6 orders vs. CPU Gibbs ≈100× vs. GPU/TPU (flips/s & energy) 6-FPGA UCSB: ~50k p-bits (async links) Synthesized 1M p-bits on 18× VP1902 (Siemens) 4,264 p-bits / ≈30k params (DBM) Multi-FPGA system to 1M+ nodes

Spotlight

Sparse Ising machine on FPGA — Sparse Ising machine — *Nature Electronics* (2022)

Hardware-trained deep Boltzmann machines — Hardware-trained DBMs — *Nature Electronics* (2024)

Higher-order Ising machines and dense behavior — Higher-order IMs & all-to-all approach — *Nature Comm.* (2024)

UCSB multi-FPGA p-computer (rack) — UCSB multi-FPGA p-computer (~50k p-bits) — lab setup

Methods I use

Massively parallel (graph-colored) Gibbs
Simulated Annealing (SA)
Adaptive Parallel Tempering (APT)
Simulated Quantum Annealing (SQA)
Isoenergetic Cluster Moves (ICM)
Non-equilibrium Monte Carlo (NMC)
Higher-order p-bits / couplers
DBM training (hardware-aware)
NQS sampling & training

Current directions

Million-node p-computers: partitioning and interconnects that preserve solution quality at extreme scale.
Probabilistic AI & NQS at scale: compact, hardware-efficient learning for scientific/quantum data.
Heterogeneous p-computers: CMOS + sMTJs to keep improving flips/J and density.

Full publication list: Publications