Memristive tabular variational autoencoder for compression of analog data in high energy physics

Feb 17, 2026 by Rajat Gupta, Y. Elangovan, Tae Min Hong, J. Ignowski, J. Moon, Aishwarya Natarajan, Stephen T. Roche, Luca Buonanno

We cram a trained variational autoencoder for calorimeter hits into a memristor ACAM by distilling the encoder into decision-tree tables, letting the in-memory hardware squeeze 48 analog channels into a 4x smaller latent stream on the fly; it runs at 330M compressions/s with ~24 ns latency and nJ-level energy, so you can ship high-rate HEP analog data off-detector without a ton of frontend silicon.

source S2