Time-Domain Computing

last updated 2026-07-27
Compute & LogicCompute & LogicConventional LogicConventional LogicBeyond Cmos LogicBeyond Cmos LogicCompute in MemoryCompute in MemoryTime-Doma…

Computing where the value of a datum is the length of a time interval — typically the delay between two pulses — rather than a voltage level interpreted as a binary digit. Arithmetic is done by delaying, racing and combining edges. The family includes race logic, unary and stochastic computing, and pulse-interval encoding.

Why anyone bothers

The pitch is always the same trade: a time-encoded operation needs far fewer transistors than its binary equivalent, because the “computation” is largely propagation rather than switching. That buys three claimed advantages:

  1. Transistor-count and energy reduction per operation — less switching activity, so lower power per useful operation.
  2. Node independence — if computation depends on propagation time rather than density, performance does not require sub-5nm nodes. This is the sovereignty and cost angle: mature-node fabs, shorter qualification cycles, a much larger addressable foundry base.
  3. Memory pressure relief — schemes that operate directly on compressed representations claim large reductions in intermediate memory traffic, which is where inference actually spends its energy.

Why it stays niche

Time is a serial resource. Classical unary and stochastic schemes buy area by spending latency, and the cost grows with operand dynamic range, so a scheme that looks brilliant on a dot product can lose badly on a full workload. Conversion between time-encoded and conventional domains is not free. And the software problem is severe: an architecture with no toolchain is a research result, not a product — the Compute Specialisation Equilibrium thesis holds that the binding brake on new silicon is mask/NRE capital and workload churn, not porting labour, but a substrate this unusual pays both.

Evaluation test

The claim to interrogate is never the ratio, it is the baseline and the workload. A transistor-count or memory-reduction number quoted against “a GPU” on a dot product tells you almost nothing about throughput per mm² per watt on real inference. Ask for the node, the die, the precision, the baseline part, and which figures are measured on silicon rather than simulated.

Related concepts

Frontier questions