Hybrid Digital-Exponent / Analog-Significand Memory Cells with Column-Shared Range Normalisation (v3 — consolidated with symmetric tunnelling SAR programming)

Abstract

This technical note is a defensive publication. It is published to establish prior art and dedicate its subject matter to the public; the author asserts no patent rights over anything disclosed and does not intend to file any patent application covering it. A non-volatile memory cell is described in which a numerical value is stored as a physically partitioned floating-point number: a small digital field holds the exponent, while a single analog charge-storage device holds the significand (mantissa), constrained to one radix interval. Range discrimination and gain normalisation are shared at the column level and amortised across thousands of cells, leaving the cell itself with only storage and select devices. The disclosure covers linear and logarithmic significand encodings; arbitrary radix, treated explicitly as a co-design parameter traded against the analog span the device must resolve; and the use of subthreshold device physics to perform antilogarithmic readout, so that logarithmically stored values can be accumulated linearly by current summation on a shared column line. A method for exponent-binned accumulation with dynamic-window pruning enables floating-point matrix–vector multiplication in an analog array without per-term digital alignment. Per-column ratiometric reference cells are disclosed for cancellation of common-mode retention and thermal drift. An error analysis establishes that relative precision is scale-invariant across the exponent range, while a single-bit exponent error produces a full radix-factor magnitude shift. This motivates asymmetric error protection: error correction on the digital exponent field, none on the analog significand — inverting the usual assumption that the analog half is the fragile one. Applications include high-dynamic-range image sensor readout, neuromorphic synaptic weight storage, and compute-in-memory inference. Prior art is surveyed honestly, distinguishing in-force patents from expired ones, and known limitations and open experimental questions are stated explicitly. Licensed CC0 1.0 (public domain dedication).