insights

Most of the cost overruns in custom silicon are not caused by design problems. They are caused by requirements that were reasonable to leave open at the start and expensive to close later. These are the ones that matter most when temperature is part of the specification.

A temperature profile, not a temperature

“200 °C” is a number, not a requirement. What a design needs is a profile:

• The maximum operating temperature, and how long the part sits there
• The minimum, which is often forgotten and sometimes harder — a part specified from −55 °C to +200 °C has a very different reference design from one specified from +25 °C
• The duty cycle: continuous, or excursions with a defined dwell
• The number of thermal cycles over the service life, and the ramp rate, both of which drive package and assembly choices rather than circuit choices
• The storage and transport range

Two applications with the same peak temperature and different profiles are different projects. A downhole tool that sees 175 °C for two hundred hours per run, several hundred times, is a fatigue problem. A furnace sensor that sits at 175 °C continuously for eight years is a wear-out problem. They are solved differently.

A service life, expressed in hours at temperature

Almost every high-temperature failure mechanism — electromigration, dielectric breakdown, bond wire intermetallic growth, charge loss in stored-charge memory — is thermally activated and accumulates with time. A lifetime requirement is therefore not “ten years”; it is hours at each temperature in the profile.

This single number sets the current density budget in the layout, the choice of metallisation, the memory strategy and the package. It is very cheap to state at the beginning and very expensive to change after tape-out.

Accuracy across the range, not at room temperature

The most common specification gap we see is a set of accuracy numbers with no temperature attached. Offset, gain error, drift, noise and timing all have temperature coefficients, and at 200 °C those coefficients dominate the error budget.

The useful way to state this is to separate what must hold across the whole range from what may be trimmed or calibrated at a known point, and to say whether calibration happens once at manufacture or periodically in the field. Those are different architectures. A part that is calibrated in the field can trade absolute accuracy for stability; a part that is calibrated once cannot.

The standard you are qualifying against

If the system has to be certified, name the standard early. Functional safety, hazardous area classification, automotive qualification and aerospace requirements each impose structure on the design — diagnostic coverage, fault detection, documentation, sometimes architectural redundancy.

Adding these after the architecture exists usually means changing the architecture. Knowing them at the start often costs very little, because the structures they require can be designed in rather than bolted on.

If you do not yet know which standard applies, that is a legitimate answer and worth saying out loud. It is better than a guess that gets designed around.

Interfaces and the rest of the board

A custom IC is defined as much by what it connects to as by what it does:

• Supply rails available, and their tolerance and noise at temperature
• The sensors or actuators at the other end, including their impedance and lead configuration
• The digital interface and what sits on the other side of it
• Cable lengths and parasitics, which are part of the measurement whether or not they are in the specification
• Board area, package constraints and any height limit

The last one matters more than it sounds. Package options narrow sharply at high temperature, and a mechanical constraint discovered late can invalidate an otherwise finished design.

How the part will be tested

Test is not a phase that happens after design. It is a set of requirements that shapes the design.

Ask early: what has to be measurable in production, and how? What has to be observable during system bring-up? Is there a self-test requirement, and does it need to run in the field or only at power-up? What does the failure of the part look like from the system’s point of view, and does the system need to detect it?

A part that is easy to characterise is a part whose numbers you can trust. Test access designed in at the start costs area. Test access added later costs a redesign.

What “done” means

Agree the acceptance criteria before the work starts:

• Which parameters are guaranteed by test, which are guaranteed by characterisation, and which are typical values
• The sample size and conditions behind the characterisation data
• What the qualification package contains
• What happens on a respin, and who carries it

This conversation is uncomfortable at the beginning and much worse at the end. A supplier who wants to have it early is telling you something useful about how they work.

What good looks like on the first call

You do not need all of the above to start a conversation. In practice the most productive first discussions cover four things:

1. What the system does, and where it sits
2. The temperature profile and the service life
3. The measurement or function, and what accuracy has to hold across the range
4. Whether a standard applies, and which

That is usually enough to say whether the problem is solvable, roughly what it involves, and what would have to be settled before anyone commits to a schedule.

If you have most of that, we can give you a useful answer quickly. If you have some of it, we can help work out the rest — that conversation is part of the job, not a prerequisite for it.