芯片验证中位与位的复杂度：从生物学借来的方法论

By Vidya Chandran Darbari and Ashish Darbari

A recent conversation about carbon intelligence and silicon intelligence stayed with me. What interested me was not which one wins, but the complexity we take on as we ask silicon to sense, interpret, decide and act.

My work as an experimental structural biologist taught me to ask what a structure could explain, and what it could not. Solving a molecular structure was not the same as explaining the behaviour of the system to which it belonged. The next question was always about mechanism: what happens when the parts interact? That question shaped my perspective as a deep generalist in life sciences, connecting molecular mechanisms, biological networks and system behaviour.

Now, as Chief Strategy and Operations Officer at Axiomise, I return to that question in a different setting.

A chip is not a cell. But investigating biological complexity can sharpen how we investigate engineered systems—provided the comparison changes a question we ask, an observation we make or an obligation we verify. Otherwise, it remains an interesting analogy.

A map is not a mechanism

In life sciences, a protein interaction map is where questions begin. A connection identifies a possible interaction; it does not establish when that interaction matters. Concentration, localization, binding affinity, feedback and timing influence the outcome.

The same messenger can stimulate or suppress an immune response depending on the receptor and the receiving cell’s state. The molecule has not become inconsistent. Our description was incomplete without its context.

Silicon engineers already understand dependencies. The useful transfer is not simply to think in networks, but to distinguish connectivity from the conditions under which connected components produce a particular behaviour.

A link can transport packets correctly, a controller can enter its power states correctly, and an interconnect can arbitrate correctly under its assumptions. Their combination can still violate a requirement if those assumptions cease to hold.

The engineering consequence is to make context explicit: which neighbouring states, available resources and event orderings does each guarantee depend on? Those conditions belong in verification, not just in integration notes.

Timing is part of the meaning

Our senses make context tangible. Combining cues helps only when their timing supports treating them as evidence of the same event. More signals do not automatically resolve ambiguity.

Consider an illustrative edge controller combining camera and distance-sensor data. Both measurements could be accurately encoded, transported without corruption and individually plausible. If they describe different moments but are treated as simultaneous, the fused estimate can represent a situation that never existed.

The bits arrive intact while the interpretation is wrong.

The biological perspective directs attention to the signal’s relationship with the state in which it is received. The engineering response is concrete: timestamp association, freshness, allowable skew and missing observations belong in the requirements.

Similar questions arise during reset, link retraining and power transitions. What happens to a transaction accepted before the transition? Which state survives recovery? When can a receiver safely treat new data as valid?

An interface is a relationship over time, not merely a set of wires. Verification must capture that relationship, particularly when the resulting interpretation controls an action.

Different uncertainties need different evidence

It is tempting to describe biology as grey and silicon as black and white. That contrast obscures more than it explains.

Biological models can combine continuous quantities, discrete events, deterministic equations and stochastic processes. These are separate distinctions. A graded response need not be random.

A digital block can have a uniquely defined next state for a given state and input while leaving a system question unresolved: which input sequences and environmental behaviours must we consider? Formal verification can examine behaviours admitted by a model under stated assumptions. That is not physical randomness.

Beneath the abstraction sits physics. Metastability is not digital logic becoming analog because an architecture grew complicated. It is physical behaviour the design must accommodate.

The differences matter because a physical measurement, statistical association and logical proof answer different questions. Learning from biology should make us more precise about where uncertainty resides—not less demanding about correctness.

Investigate what would distinguish the explanations

My life-sciences training made me wary of assuming that a symptom identifies its cause.

Consider an illustrative chiplet system, not a customer incident. A transaction times out. A power controller has gated a clock before outstanding traffic drained. It will restore the clock after receiving a wake request, but that request is queued behind the stalled traffic.

Each side waits for the other. The timeout is the symptom; the mechanism is a circular dependency. Extending the timeout would not remove it.

Other mechanisms could produce the same symptom: a lost credit, inconsistent reset state or starvation. Reproducing a trace does not distinguish those explanations.

Here, experimental reasoning suggests a useful question: which intervention would separate the hypotheses? What happens if the clock remains running? What if wake delivery is independent of normal traffic? Does the failure persist when downstream service changes?

These are familiar verification activities. Their explanatory purpose matters: choose an investigation because its outcome challenges a mechanism, not merely because it produces another passing run.

A counterexample also needs interpretation. It may expose a defect, an unrealistic environment or behaviour introduced by abstraction. A passing result needs equal scrutiny: did the assumptions exclude the situation we needed to examine?

Connect evidence across layers

Life sciences investigates activity at different levels: genome sequence, gene expression, protein machinery and metabolism. These measurements are not interchangeable. Discrepancies can reveal regulation, compensation or an incomplete explanation.

There is no exact mapping from these layers to a chip. The transferable principle is to connect evidence across levels rather than expect one measurement to explain everything.

RTL and configuration describe one view. Workload activity, transaction traces and physical telemetry provide others. Imagine a workload increasing power demand, a thermal policy reducing frequency and a controller approaching its deadline. Neither the transaction trace nor the temperature reading alone establishes whether the complete sequence was permitted.

I see an opportunity for tools that organize observations around shared events and competing explanations. AI-assisted analysis could help propose relationships, but correlation should lead to a testable hypothesis, not be labeled a root cause.

Observability should follow the same reasoning. For the clock-gating example, we need to relate outstanding traffic, clock decisions, wake generation and wake delivery. More logs are not necessarily better evidence if those relationships remain invisible.

Turn mechanisms into obligations

Once we understand a mechanism, we should not leave it in a debug report.

What must prevent the circular dependency? Must traffic drain before clock gating? Must wake delivery remain independent? What happens if the normal sequence is interrupted?

These questions turn explanation into architectural obligations. This is the connection I see between biological reasoning and formal-first engineering: one directs attention towards context and competing mechanisms; the other helps express and check precise requirements in a design model.

At Axiomise, formal-first means asking those questions while the architecture can still change. Credits must not be lost or double-counted. Ordering must be preserved where required. Reset and recovery must restore compatible endpoint states.

Progress needs its own argument. Nothing being corrupted does not establish that an accepted request eventually completes. Service and fairness assumptions must be justified by the architecture.

Local proofs need the same treatment. If one component assumes a neighbour responds, where is that response established? Individually successful checks cannot establish a system claim when their assumptions are incompatible or circular.

Formal-first is not formal-only. Simulation, emulation, physical analysis and system validation contribute different evidence. The task is to connect them around identifiable questions.

Ownership should follow the obligation

If a failure mechanism crosses power management, traffic control and recovery, preventing it cannot depend solely on each team completing local checks.

Someone must own the cross-system obligation: what must remain true, which components contribute, which assumptions are required and which changes trigger reassessment. A revised power policy should reopen the progress argument if it changes its conditions.

This does not diminish specialist expertise. It gives the relationships between specialisms the same standing as the components they produce.

Other fields will not hand silicon engineering a ready-made solution. They can help us recognize incomplete descriptions and design better investigations. Biology’s contribution is not permission to accept ambiguity, but a discipline for challenging explanations.

Looking back at structural biology, I recognize the temptation to let a beautifully resolved model feel like the end of the investigation. The important question was what it helped explain—and what should follow.

The bits can be binary. Our understanding of the system is not automatically complete.

Ashish Darbari is the founder and CEO of Axiomise.