MXFormer: A Microscaling Floating-Point Charge-Trap Transistor Compute-in-Memory Transformer Accelerator

Feb 12, 2026 by G. Karfakis, S. Chakrabarty, V. K. Jacob, S. Qiao, Subramanian Iyer, S. Pamarti, Puneet Gupta

Built MXFormer, a hybrid weight-stationary CIM accelerator that packs ultra-dense Charge-Trap Transistor MXFP4 arrays to keep hundreds of millions of Transformer weights on‑chip and ditch weight movement, combining analog CTT MLPs with precise digital attention to get near-digital accuracy while blasting throughput and efficiency for short-sequence inference. If you care about pushing compute and resident weight density without retraining, this shows how microscaled floating-point charge‑trap arrays and a statically pipelined FWS datapath beat state-of-the-art digital, hybrid, photonic and FWS designs by large margins.

source S2