Research / Associative-memory architectures

IEEE TCAS-I 2026 · First and Corresponding Author

CAMformer: Binary Associative Memory Is All You Need

CAMformer replaces much of transformer attention's dense arithmetic and score movement with parallel associative-memory retrieval.

CAMformer maps binary transformer attention onto a voltage-domain content-addressable memory array
Binary queries are compared with stored keys in parallel, then a hierarchical top-k pipeline selects values for contextualization.

9,045

queries per mJ, a reported 10× attention-layer energy-efficiency gain

191

queries per ms, a reported up to 4× attention-layer throughput gain

0.26 mm2

single-core area, reported as 6–8× smaller

1.12%

mean PVT sensing error, reported as 7× lower

Problem

Transformer attention compares every query with every key. The arithmetic, intermediate scores, and data movement therefore grow quadratically with sequence length. CAMformer asks whether attention can instead be implemented as a memory lookup whose physical operation is similarity search.

Approach

My Role

As first and corresponding author, I led the cross-layer design of CAMformer, from voltage-domain associative sensing through the sparse attention pipeline and system evaluation.

Citation

T. Molom-Ochir, B. F. Morris, M. Horton, C. Wei, C. Guo, B. Taylor, P. Liu, S. X. Wang, D. Fan, H. Li, and Y. Chen, "CAMformer: Binary Associative Memory Is All You Need," IEEE Transactions on Circuits and Systems I: Regular Papers, pp. 1–14, 2026. DOI