Architecture
Accuracy that does not depend on the clock
Analog-to-digital conversion is the backbone of modern electronics. Every traditional architecture requires good clocks, and ours doesn't.
What a converter actually does
An analog-to-digital converter takes rapid measurements of a continuously varying analog signal and reports them as digital numbers. The Shannon–Nyquist theorem tells us how often those measurements have to happen, and requires they happen at perfectly even intervals.
When a measurement lands early or late the recorded picture blurs and information is lost. The industry's answer has been to build better clocks. That takes a lot of power and engineering.
Why every conventional design hits the same wall
Conventional ADCs have many different architectures, but they all perform Shannon-Nyquist sampling. Whether they are successive approximation, flash, or delta-sigma, or hybrid or pipelined, all of them depend completely on a stable, high-precision clock to measure exactly when expected. Whatever the internal architecture, any deviation in timing creates errors that propagate through the system and limit the amount of information that can be captured.
When ADCs are used to receive signals in networks, the information lost from timing imprecision gets more pronounced with more complicated signals. Signals with more amplitude levels, or more options for each symbol, carry more information per symbol.
Signalling protocols used to use two levels, but today the fastest ones use 4 levels. This is referred to as PAM4. There is a lot of interest in increasing the modulation even higher, to 6 or 8 levels per symbol, but today's clocks can't strike precisely enough to support Shannon-Nyquist sampling of high modulation signals. That timing dependency is what closes off higher data rates on traditional systems. We handle worse timing better.
Two stages instead of one
Rather than translating the waveform directly with amplitude measurements, we separate capture from interpretation. Capturing only and exactly the information needed to produce analytically tractable readings is a very hard problem. We split it into two easier problems.
Stage one: Capture
Coarse measurements, sampled in parallel
A bank of ordinary bandpass filters decomposes the signal into frequency channels. Each channel is sampled separately and slowly, below the Nyquist rate. No single measurement has to carry the whole signal, so no single measurement has to be perfect.
Stage two: Interpret
A network that learns its own circuit
A lightweight neural network approximates the inverse transfer function of the analog front end, mapping those channels back to the original signal and correcting for the distortions the hardware introduced. We don't use complicated analog calibration mechanisms because we don't need the measured value to match some true value. The interpretation is the calibration: there is no separate calibration mechanism.
What follows from the architecture
- Ordinary parts are good enough
- Our Phase I filter array used standard 5% tolerance passives, deliberately not precision-matched. The fabricated filters deviated substantially from their simulated models, with shifted cutoffs, broader roll-off, and responses we had not designed for. Because the network is trained on the circuit it actually has, rather than the circuit we drew, it adapted without intervention.
- The full signal band is preserved
- Unlike other sub-Nyquist approaches such as Compressed Sensing, we make no assumptions about the signal before it reaches the device. The neural network learns the behaviour of its analog front end circuitry, not the content of what passes through it, so anything within the frequency band is recoverable.
- Redundancy becomes yield
- At small node sizes a single misbehaving block on a critical path can ruin a die. Capturing overlapping, partial information across many channels means no single channel is critical, and the network turns that redundancy into tolerance for manufacturing variation.
Simulated, then measured
The claims on this page are supported by hardware measurements from two generations of physical prototype.