The mechanism
In analog in-memory computing, stored conductance can encode a weight and an applied voltage can encode an input. Currents combine to form a multiply–accumulate result that must be sensed and often digitized. IBM has demonstrated mixed-signal phase-change-memory research chips using this principle. Photonic tensor-core research instead encodes and combines signals using optical components, with electronic interfaces still required around the computation. Both approaches can reduce certain data movements or exploit physical parallelism, but noise, drift, calibration, conversion and unsupported operations determine complete-system usefulness. Research demonstrations establish a mechanism and measured experiment, not automatic production readiness for large AI services.
- Choose how numerical values are represented by conductance, voltage or an optical signal.
- Program weights and calibrate the physical response, including device variation and drift.
- Perform the physical combination that approximates the desired matrix operation.
- Sense and convert the result, accounting for noise and the precision required by the next stage.
- Evaluate model quality and total energy including peripherals, communication and operations performed outside the analog or optical core.
Science and assumptions
Ideal conductance accumulation
In the ideal conductance-array model, each output current sums input voltage times its corresponding conductance.
This electrical conductance equation models an analog crossbar mechanism; it is not a universal equation for every photonic architecture.
Symbols and units
- Iⱼ
- Output current at column j
A
- Gᵢⱼ
- Programmed conductance from input i to output j
S (siemens)
- Vᵢ
- Applied input voltage
V
- i,j
- Input and output indices
integer indices
Gij is conductance in siemens, Vi voltage in volts and Ij current in amperes; the sum represents an ideal linear dot product.
Assumptions Linear devices, negligible wire drops and ideal readout; real arrays require calibration and may use differential encoding.
Engineering example
Synthetic error budget: if sensing and weight drift each consume part of a permissible error margin, a larger array may need recalibration or wider digital correction even when its core operation is fast.
Distinctions and tradeoffs
Light does not make data movement or heat disappear, and an analog multiply–accumulate does not supply arbitrary exact digital arithmetic.
Physical parallelism may reduce core arithmetic energy, while precision restoration, calibration and data conversion can consume much of the apparent benefit.
Inference tolerance to numerical error creates opportunities, but model quality and full operator coverage must survive real device variations.
Specifications and interfaces
- Evidence status
- Research demonstration
Laboratory arithmetic is separate from a qualified deployed service.
- Accounting boundary
- Array plus converters, control and memory
Laser or conversion energy must be counted where the architecture requires it.
Interfaces
- Encoded inputs and weights Input
- Voltages/conductances or optical signals
- Readout and digital control Output
- Digitized results, calibration feedback and subsequent operations
Check your understanding
What belongs in a fair analog-versus-digital energy comparison?
- Converters, control, memory traffic and achieved model quality
- Only the crossbar’s current
- Only the optical propagation time
Answer and reasoning
Converters, control, memory traffic and achieved model quality
The complete task needs interfaces and acceptable outputs; excluding them can turn a component result into an unsupported system claim.
Evidence and scope
The IBM and Nature sources describe particular research systems. No commercial deployment scale, universal energy advantage or precision equivalence is asserted.
Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism.