Why Probabilistic Computing?¶

Imagine you need to multiply two numbers that each lie between 0 and 1 — say image pixel intensities, neural activations, or channel soft-bits. In classical binary arithmetic you build a wide multiplier array. In stochastic computing (SC) you instead encode each value as a random bitstream whose density of 1s equals that value, then multiply with a single AND gate.

graph LR
    A["Value A = 0.75<br/>bitstream ~11101110…"] --> AND["AND gate"]
    B["Value B = 0.50<br/>bitstream ~10101010…"] --> AND
    AND --> C["Product ≈ 0.375<br/>density of 1s"]

The catch: you wait \(N\) clock cycles for a good estimate, and the answer is noisy. When the application already tolerates noise — neural nets, vision, iterative decoding — that trade can win on area and energy.

The energy intuition¶

CMOS energy is paid per toggled transistor. A 16×16 binary multiplier toggles thousands of gates every cycle. An SC multiplier toggles one gate per cycle for \(N\) cycles. If \(N\) is modest (tens to hundreds) and the rest of the SoC is memory-bound or already approximate, SC can be attractive.

\[E_{\text{SC-mult}} \approx N \cdot E_{\text{AND}} \quad\text{vs}\quad E_{\text{bin-mult}} \approx E_{\text{array}}\]

SC also pairs naturally with compute-in-memory and unary datapaths where values are already pulse-density coded.

What you get — and what you pay¶

+------------------------------------------------------------------+
|  WIN: tiny arithmetic · soft-error softening · easy nonlinearity |
|  PAY: latency (N cycles) · RNG / SNG cost · correlation bugs     |
+------------------------------------------------------------------+

Property	Binary exact	Stochastic / probabilistic
Multiply (unipolar)	Large array	`AND` (if independent)
Add (scaled)	Adder tree	MUX / OR / accumulator hybrids
Precision	Bit-width \(W\)	Stream length \(N\) (and encoding)
Errors	Hard faults hurt MSB	Many flips → small density change

Interactive lab: bitstream multiply¶

Drag the two probabilities and watch an AND-based unipolar multiplier estimate the product from a finite stream:

Where to go next¶

Bitstream Encoding Made Easy — unipolar vs bipolar
Correlation & Independence — why shared LFSRs break products
Hardware Taxonomies — circuit families with DOI citations

Literature without PDFs

Browse curated titles and DOIs on the PC Paper Map. This portal never hosts publisher PDFs.