Honors Thesis · Brain-Inspired Computing Lab · In progress
Reservoir Computing Hardware on ESO Memristors
Readout and driver circuitry for a reservoir computer built on entropy-stabilized oxide (ESO) memristors, targeting TSMC 65nm. Hardware limits are modeled in simulation before any silicon is committed.
Spring 2026 – present
- ADC spec set by sweeps
- 3-bit
- Audio-MNIST accuracy
- ~98.5%
- real devices validated against
- 16
- TSMC target node
- 65nm
The problem
Reservoir computing uses the physical dynamics of a device, here ESO memristors, as the “hidden layer” of a neural network. Only a simple readout has to be trained. The catch is that the surrounding circuitry decides whether the idea survives contact with silicon. Readout amplifiers, ADCs and drivers all add noise, quantization and voltage limits.
Approach
- Hardware-aware simulation. I built a PyTorch model of the full signal chain: memristor reservoir → transimpedance amplifier (TIA) readout → ADC. The model includes quantization, ENOB noise and clamped supply rails. Large batch sweeps ran through Slurm on OSU’s H100 GPU cluster.
- Spec-driven design. I swept ADC bit-width and ADC sharing to find the cheapest converter that preserves accuracy. The result is a 3-bit ADC specification that keeps ~98.5% accuracy on Audio-MNIST spoken-digit classification.
- Validation against silicon data. I found and fixed readout-training bugs in the group’s existing code. I then validated the simulations against measured data from 16 real devices.
Current phase: analog design
I’m writing specifications for the input driver (DAC) and designing it in Cadence Virtuoso. It uses tri-state row drivers with 1.2 V program / 0.3 V read pulses. This keeps everything within the 65nm 1.2 V limit, so no level shifters are needed. RTL and Verilog-A device modeling come next.
What’s next
- Thesis defense, March 2027
- Targeting a MWSCAS 2027 publication
Supervised by Sangmin Yoo, PhD. Process design kit details stay under NDA, so this page describes the design at the specification level only.