🧠 Module 3 · From Biology to Algorithm · Chapter 3.7 · 13 min read

Memristor ↔ Synapse Mapping

Same math, different substrate — SIDRA's central thesis.

Prerequisites

What you'll learn here

Write the one-to-one mapping between synaptic weight w and memristor conductance G
Detail the parallel between LTP/LTD biology and SET/RESET memristor physics
Discuss where the memristor wins (speed, control) and loses (energy, plasticity diversity)
Compare the 256-level SIDRA cell information content with real neural-network weights
View an entire crossbar as a synaptic weight matrix and understand it as 'synapse + neuron compression'

Hook: SIDRA's Central Thesis

Module 2 covered the chemistry of the HfO₂ memristor: oxygen-vacancy filament, SET/RESET, 256 discrete conductance levels. The first six chapters of Module 3 unpacked the biological synapse part by part: AMPA receptors, LTP/LTD, Hebbian, BCM.

Now bridge them. SIDRA’s central thesis in one sentence:

The HfO₂ memristor plays the same mathematical role as a biological synapse — weight × input = output. The same plasticity rules (LTP=SET, LTD=RESET) operate on a different substrate.

This mapping isn’t accidental — the physics is the same: ion motion within a lattice → conductance change. In HfO₂ it’s O²⁻ vacancies; in a synapse it’s AMPA receptors. The mechanism differs; the math is one.

If the thesis holds: SIDRA can build brain-inspired AI infrastructure. If not: SIDRA is “just an efficient matrix multiplier.” This chapter audits the thesis and shows how far it carries.

Intuition: A Side-by-Side Comparison

Synapse and memristor, on one page:

Property	Biological synapse	HfO₂ memristor (SIDRA)
Weight	$w$ (AMPA receptor count)	$G$ (conductance, filament thickness)
Multiply rule	$I_{\text{post}} = w \cdot V_{\text{pre}}$	$I = G \cdot V$ (Ohm)
Sum (integration)	Postsynaptic EPSP summation	KCL: column current = $\sum V_i G_i$
Strengthening (LTP)	NMDA Ca²⁺ → AMPA insertion	Positive voltage pulse → SET (filament grows)
Weakening (LTD)	Low Ca²⁺ prolonged → AMPA endocytosis	Negative voltage pulse → RESET (filament breaks)
Plasticity timescale	minutes (LTP), hours (long-term)	nanoseconds (SET), persistent
Information content	~5 bits (16-64 discrete levels, est.)	8 bits (256 levels, guaranteed)
Cell size	10-20 nm (cleft)	100 nm (Y1), 28 nm target Y10
Per-event energy	~10 fJ (synaptic event)	~0.1 pJ (memristor read), ~10 pJ (SET)
Lifetime (cycles)	Complex (year-scale retention)	10⁶-10⁹ SET/RESET
Native noise	High (vesicle stochastic)	Low (deterministic, but quantization noise)

Important: the comparison is rough; real synapses are much more complex (NMDA gating, dendritic non-linearities, glial modulation). Memristors aren’t simple either (drift, retention, endurance). But the shared mathematical core is the bridge’s load-bearing pillar.

Formalism: The Mapping in Detail

L1 · Başlangıç

Common equation:

Both synapse and memristor are single-input, single-output linear multipliers:

\text{output} = \text{weight} \times \text{input}

Synapse: $I_{\text{postsynaptic}} = w \cdot R_{\text{presynaptic}}$ ( $R$ = presynaptic firing rate, or single-event indicator).

Memristor: $I = G \cdot V$ (Ohm).

Multi-input summation:

In the brain, a neuron sums currents from 1000+ synapses (postsynaptic membrane). KCL:

I_{\text{neuron}} = \sum_{i=1}^{N} w_i \cdot R_i

In a memristor crossbar, 256 cells share a column. KCL:

I_{\text{column}} = \sum_{i=1}^{256} G_i \cdot V_i

Same equation, different symbols. This is the mathematical proof that a SIDRA crossbar “imitates” a neuron.

L2 · Tam

Plasticity mapping:

LTP (synapse): $\Delta w = \eta \cdot f(\text{pre-post timing})$ → AMPA insertion.

SET (memristor): $\Delta G = g(\text{voltage pulse width, amplitude})$ → filament grows.

Mechanism similarity:

Synapse LTP: Ca²⁺ influx → CaMKII phosphorylation → AMPA → more channels → effective $G$ rises.
Memristor SET: O²⁻ vacancy motion → filament grows → metallic bridge → $G$ rises.

Both:

Require ion motion
Require above-threshold voltage (NMDA activation in synapse, V_set in memristor)
Are reversible but asymmetric (LTD/RESET use a different mechanism)
Are thermally activated (Arrhenius dependence, retention)

Plasticity speed:

Event	Synapse time	Memristor time
Single-event	1-5 ms (EPSP)	1-10 ns (read), 10-100 ns (SET)
LTP onset	30 s-min	1 SET pulse (10-100 ns)
Persistence	hours-years	years (retention 10⁶-10¹⁰ s @ 85°C)

The memristor performs plasticity 10⁶× faster — both an advantage (fast training) and a disadvantage (doesn’t scale to biological dynamics). For SIDRA the advantage dominates: synaptic weight updates at the nanosecond scale.

Information content match:

Synapse: log₂(N_levels). Bender et al. (2007) estimated 4-6 bits, but precise measurement is hard. Modern estimates fall in the 1-7 bit range.

Memristor SIDRA Y1: 256 levels = a guaranteed 8 bits (during programming). Post-drift, perhaps 6-7 effective.

SIDRA edge: memristor delivers a guaranteed 8 bits; synapse estimates are ~5. The memristor carries ~60% more information than a biological synapse — silicon’s analog advantage.

L3 · Derin

Limits of the mapping (important caveats):

Synapses aren’t linear: EPSP summation is linear at small amplitudes but saturates near threshold. The memristor is Ohmic (linear). This means the memristor is more linear than a synapse at small inputs.
Synapses are NMDA-gated: only when pre + post are simultaneously active does Ca²⁺ flow → learning. The memristor responds to every voltage pulse. So the synapse has a “smart gate”; the memristor doesn’t. SIDRA adds this with the surrounding circuit (CMOS gate).
Synapses are stochastic: vesicle release is probabilistic (p = 0.1-0.9). Memristors are deterministic (programmable). The brain leverages noise (regularizer); the memristor loses that. SIDRA Y10 target: controlled-stochastic memristor (noise added by design).
Synapses “always run”: continuous energy (ATP pump). Memristors are non-volatile (no energy after programming). Memristors win here — zero idle power.
Synapses “grow/shrink” during learning: AMPA insertion, dendritic spine growth. Memristors don’t change physical dimensions (only internal filament configuration). Synapses are alive; memristors are frozen.
Multi-timescale plasticity: the brain runs STDP (ms) + LTP (s) + synaptic scaling (hours) + structural plasticity (days-months) in parallel. Memristors are mostly single-timescale. SIDRA Y100 target: multi-timescale plasticity implementation (multi-bit + slow drift + fast STDP).

How far the SIDRA thesis carries:

✅ Math (MVM): identical.
✅ Plasticity general form (LTP/LTD ↔ SET/RESET): same.
✅ Hebbian learning (V_pre · V_post · Δt): natural.
⚠️ Plasticity diversity: SIDRA is single-timescale (until Y100).
⚠️ Stochasticity: SIDRA is deterministic (until Y10).
❌ NMDA-gated coincidence detector: SIDRA emulates with surrounding circuitry, not the device itself.
❌ Structural plasticity (forming new connections): SIDRA hardware is fixed; the brain is flexible.

Bottom line: SIDRA reproduces ~70-80% of biological synaptic function. The remaining 20-30% comes post-Y100 with new devices (organic PEDOT:PSS, ferroelectric, FeFET).

Experiment: A SIDRA Crossbar = Synaptic Weight Matrix

A 256×256 SIDRA crossbar = 65,536 memristors = roughly the synapses of a single neuron (cortical pyramidal neurons have 5,000-50,000 synapses).

Scenario: building a single “artificial neuron”. 256 inputs + 1 output, with weights.

How many memristors per neuron?

256 synapses × 1 weight = 256 memristors (one crossbar row).

How many neurons fit in a single 256-input MLP layer?

256 rows × 256 columns = 256 neurons.
I.e., the crossbar = a 256-input, 256-output MLP layer.

SIDRA Y1 (419M memristors): 419M / (256×256) = 6400 crossbars. Each = a 256-input, 256-output connection block.

Equivalent biological neural net:

6400 × 256 = 1.64M output neurons, each with 256 synapses.
That’s about 1640× a cortical mini-column (V1 visual cortex mini-column ≈ 1000 neurons).
Or: ~1% of the connections of all of V1 visual cortex (~140M neurons).

Brain comparison:

A pyramidal neuron has 10,000 synapses. A SIDRA row settles for 256.
Fully-connected: SIDRA averages 256 synapses/cell vs the brain’s 10,000. The brain is 40× more connected.
But SIDRA is faster (10⁶× plasticity), more precise (8-bit guaranteed).

Bottom line: SIDRA Y1 can model a small biological network (~1.6M post-neurons, limited connectivity) — enough for MNIST classification or small language model inference. True brain scale is on the Y10-Y100 horizon.

Quick Quiz

1/6What is the memristor analog of synaptic weight w?

Lab Exercise

Can SIDRA Y1 actually simulate “a person-sized chunk of brain”?

Data:

SIDRA Y1: 419M memristors.
C. elegans (roundworm) full synaptic connectome: ~7000 synapses, 302 neurons. (Connectome published in 1986.)
Drosophila (fruit fly) brain: ~3 × 10⁵ neurons, ~10⁸ synapses.
Human cortical mini-column: ~10⁴ neurons, ~10⁷ synapses.

Questions:

(a) C. elegans full connectome uses what fraction of SIDRA Y1? (b) Drosophila brain: how many Y1 chips? (c) Human cortical mini-column: how many Y1 chips? (d) Y10 target = 10B memristors. Which biological network does it fully match? (e) Y100 target = 100B memristors. Which biological network?

Solutions

(a) 7000 / 419M = 0.0017%. Y1 has the capacity of ~245,000 C. elegans worms in parallel. Sounds outrageous, but modern AI models are far more complex than C. elegans.

(b) 10⁸ / 4.19×10⁸ = 0.24 → a Drosophila alone doesn’t fill Y1. About 4 Y1 = 1 fly brain. The Drosophila full connectome was completed in 2024 (FlyWire); the fly is staggeringly complex (flight, olfaction, mating behavior).

(c) 10⁷ / 4.19×10⁸ = 2.4%. Y1 = ~40 cortical mini-columns (the human brain has ~10⁵ mini-columns; this is ~0.04% of them).

(d) Y10 = 10B memristors = 24× Y1. All of Drosophila brain. Cortical area: 10⁹ synapses ≈ a small visual cortex (1% of V1).

(e) Y100 = 100B memristors. A full cortical mini-macro column, or all of Drosophila + a large visual cortex chunk. Still 0.1% of the human brain (~10¹⁴ synapses).

Lesson: SIDRA’s roadmap isn’t to imitate biology — it’s to provide useful AI function. Simulating C. elegans is interesting science; Y1’s actual goal is MNIST inference + edge AI, far more aggressive use. Hardware capacity gap ≠ AI capability gap — making a model smart is a different problem from hardware size.

Cheat Sheet

Central thesis: $w_{\text{synapse}} \leftrightarrow G_{\text{memristor}}$ . Same math.
MVM mapping: synaptic integration $\sum w_i R_i$ ↔ crossbar column current $\sum G_i V_i$ .
Plasticity mapping: LTP ↔ SET, LTD ↔ RESET. Different mechanism, same direction.
SIDRA wins on: speed (10⁶×), bit content (8 vs ~5), idle energy (non-volatile vs ATP), control (deterministic).
Synapse wins on: multi-timescale plasticity, structural flexibility, NMDA-gated coincidence, organic noise.
Thesis fidelity: ~70-80% of synaptic function direct; the rest (structural, gated) on the post-Y100 horizon.

Vision: Synapse + Memristor Hybrid and the Bio-Compatible Chip

As the central thesis is validated, the horizon expands:

Y1 (today): HfO₂ memristor = inorganic analog synapse. The thesis’s first practical demonstration.
Y3 (2027): Surrounding CMOS implements NMDA-gated coincidence; controlled noise added by design.
Y10 (2029): Multi-timescale plasticity — fast STDP + slow scaling on the memristor at once.
Y100 (2031+): Structural plasticity — programmable routing (FPGA-style connection decisions + memristor weights). Hardware analog of the brain “forming new synapses”.
Y1000 (long horizon): Organic PEDOT:PSS synapse — bio-compatible, brain-implantable. SIDRA’s bio-electronic generation.

Strategic opportunity for Türkiye: the synapse-memristor hybrid is a category outside the classical silicon industry. If Türkiye combines (1) ALD infrastructure (Module 2), (2) memristor research capacity (TÜBİTAK BİLGEM, universities), (3) workshop discipline → the world’s first bio-compatible neuromorphic fab is possible in Türkiye.

Unexpected future: Neuralink + SIDRA hybrid. Neuralink electrodes read brain signals → SIDRA memristor crossbar processes in real-time → returns interpreted answers to the brain. Closed-loop brain-computer interface. A child sits down once, never types, controls devices by thought. 2035-2040 horizon; Türkiye has the right infrastructure to compete in this race.

Prerequisites

What you'll learn here

🪝 Hook: SIDRA's Central Thesis

🧭 Intuition: A Side-by-Side Comparison

📐 Formalism: The Mapping in Detail

🧪 Experiment: A SIDRA Crossbar = Synaptic Weight Matrix

📝 Quick Quiz

🛠️ Lab Exercise

🗂️ Cheat Sheet

🔮 Vision: Synapse + Memristor Hybrid and the Bio-Compatible Chip

📚 Further Reading