Research

My research asks a single question: how can memory do more than store data?

I build emerging-device circuits and associative-memory architectures that execute search, retrieval, and iterative reasoning where data already lives. The program connects three layers of the stack: physical memory substrates, circuits for parallel similarity, and systems that use those primitives for modern AI workloads.

Emerging devices Associative memory Search and reasoning

Three-dimensional memristive memory arrays and efficiency comparisons — Emerging memories can combine storage, local dynamics, and computation within the same physical substrate.

01 / EMERGING DEVICES

Memristive Substrates and Circuits

Research question: Which device and circuit primitives let memory execute similarity, local dynamics, and feedback with minimal data movement?

Current evidence includes collaborative work on memristive cellular neural network hardware that combines a 65 nm CMOS network, multilevel nonvolatile memristor synapses, and a digital twin for image and partial differential equation processing.

Selected evidence: Memristive cellular neural networks, Nature Electronics (2026)

Next direction: Nonvolatile and three-dimensional memory structures that support similarity search, local interaction, and feedback without repeatedly returning data to a processor.

CAMformer maps transformer attention onto a content-addressable memory array — Associative-memory circuits compare a query against many stored keys in parallel.

02 / ASSOCIATIVE MEMORY

Associative-Memory Circuits and Architectures

Research question: How can similarity search and attention become native memory operations instead of dense arithmetic followed by data movement?

CAMformer maps binary attention to a voltage-domain CAM and a sparse top-k pipeline. MonoSparse-CAM uses tree-model sparsity and matchline monotonicity to skip redundant CAM operations. Related work studies scalable NoC-partitioned arrays and synthesizes the design space across 37 CAM cells.

Selected work: CAMformer · MonoSparse-CAM · Hamming Attention Distillation · CAM circuits survey · NP-CAM

Next direction: Programmable associative arrays that support multiple similarity operations while remaining robust to device and circuit variation.

Multi-primitive in-memory accelerator for Monte Carlo tree search — Search and reasoning require several memory operations coordinated through a feedback loop.

03 / SEARCH AND REASONING

Hardware for Inference and Reasoning

Research question: How can memory systems support irregular, multi-phase search and the feedback needed for test-time computation?

IMC-MCTS decomposes Monte Carlo tree search into five memory and logic primitives, then coordinates them as one accelerator. Complementary work maps neuro-symbolic policies to CAM lookup, connecting learned perception with efficient rule execution.

Selected work: IMC-MCTS · Efficient Neuro-Symbolic Policy

Next direction: Memory fabrics with local update and feedback that make multi-step search and test-time reasoning energy efficient.