| Episodes | 102 |
| Mentions | 137 |
| Cited here | 46 |
| First — last | #2 — #725 |
| Top guests | Henry Ott, Alan Wolke, Hank Zumbahlen |
| Related | transistor · digi key · microcontroller · pcb · analog to digital converter |
An integrated circuit (IC) is a semiconductor device in which a complete circuit function is fabricated as a single unit, sealing its internal operation inside a package that presents only a defined interface to the designer using it.[165][725] The scope of what qualifies as an integrated circuit has expanded continuously: in the 1970s and early 1980s the term covered a timer, an operational amplifier and a family of logic gates, with anything more ambitious built from discrete transistors, whereas modern practice treats the IC as the atomic unit from which entire systems are assembled.[704][165] Integration typically consolidates cost onto the silicon, making a single integrated device cheaper than the equivalent collection of separate packages, and has consequently moved most electronics engineering from circuit construction to system-level design.[253][185]
History
The first logic family intended for broad use was released in 1964 and aimed at military applications, with a commercial version following roughly two years later; the first IC-based project appeared in the hobby press in 1967, with the devices costing on the order of a pound and a half each.[207] By the 1970s and early 1980s the working definition of an integrated circuit comprised the 555 timer, the 741 operational amplifier and 7400-series logic, and designs of that period are dominated by discrete transistors precisely because anything beyond those parts had to be built from them.[704]
In the Apollo era, integrated circuits were new and unreliable, so space-grade parts were individually screened and hand-picked to achieve the required reliability, and aerospace constituted most of the market.[401] Consumer volume later inverted this practice: across production runs of tens of millions of units a month, any appreciable failure rate is financially impossible, which forced commercial-part reliability high enough that unscreened parts could be used where screened ones were once mandatory.[401] Devices were also once expensive enough to support a product category of dedicated IC testers, used to justify desoldering parts from scrap boards, testing them and returning them to stock; cheap online component supply removed the economics of that practice.[408]
The expansion of integration proceeded by migrating functions into packages as enabling components shrank. Isolation, for example, was long implemented with a transformer plus separate driver and receiver circuits; the contribution that made an integrated isolator possible was developing transformers small enough that multiple channels could fit inside an ordinary IC package.[185] Integration also drove new infrastructure: the low-speed serial bus now most widely used between devices was created explicitly for inter-integrated circuit communication, because a television set could not practically be wired with the multi-gang switches that preceded it.[396]
The cumulative effect was a change in what electronic design work consists of. Early solid-state design meant building transistor circuits out of transistors; the modern equivalent is systems engineering, assembling systems from integrated circuits whose functionality is sealed inside each package.[165] Boards carrying a hundred separate logic devices no longer appear in new design work, the functionality having moved into microcontrollers, programmable logic or purpose-built devices from major manufacturers.[597] Because an IC has already been debugged by its designer, the board-level task reduces to confirming that a signal is present and the supply is working, with everything beyond that being software.[199]
Economics of integration
Integration is usually cheaper than assembling the equivalent function from separate packages, because the cost consolidates onto the silicon: three packages meeting a specification typically cost more than one integrated device.[253] Highly integrated parts are built for a specific problem rather than for generality, and within that problem a discrete implementation rarely competes; outside the intended application, the same integration may deliver neither the accuracy nor the capability required.[348] Even complex-sounding analog parts decompose into familiar blocks—an instrumentation or difference amplifier is three operational amplifiers with precision resistors—so what a vendor sells is a packaged system built from the same components a designer would otherwise assemble.[15]
Design knowledge accumulates into silicon and the firmware around it, converting specialist capability into a commodity: an inertial measurement unit that once cost hundreds of dollars became a part costing cents, and the engineering formerly required to build one is now bought rather than performed.[601] Designing to a low price point shows in device count; a volume-engineered product can reduce to two main chips, one integrating the whole sensor function and one doing the processing.[130] At the other extreme, small single-function “jelly-bean” parts must be extremely cheap by nature—around twenty-four cents at a thousand pieces and a few cents at high volume—because an additional device cannot otherwise be justified in a design review, and the manufacturer will only engage at quantities in the hundred thousands.[96]
The choice between a pre-integrated module and the bare devices is a volume calculation: a module gives short time to market and easier development, while implementing the chips directly becomes cheaper once deployment is large.[376] Optimisation itself means different things to different designers: an analog designer may replace one integrated part with ten or twenty discrete transistors, while measured on total overhead a vendor’s complete solution costing a dollar more may be the genuinely optimised choice, depending on whether board area, unit cost or engineering time is the binding constraint.[2] The judgment therefore runs in both directions—engineers report reluctance to use highly integrated parts where a discrete implementation feels more legitimate, though modern systems could not be built at all without such devices, which is why catalogues hold hundreds of thousands of parts; equally, a fifty-cent IC should not be specified where a two-cent transistor performs the function.[88][424] Part selection should accordingly be revisited against the actual requirement rather than the assumed one, since a device chosen for higher performance may deliver capability the design does not need while a cheaper part offers the same function.[253]
Adopting a device late in its life can be a deliberate strategy: a protocol stack at version seven has absorbed years of other people’s difficulties, allowing one engineer to implement alone what previously required a team.[516] Selection is also decided by non-technical factors more often than datasheets suggest: reuse of a device already qualified across a company’s products, an engineer’s preference for parts they find easy to work with, and refusal to specify a supplier that has caused problems before.[565]
Sourcing and supply
Component shortage inverts normal purchasing practice: critical integrated circuits have been bought years before the designs that would use them existed, a poor use of capital that was nonetheless the only way to guarantee availability.[614] On a hundred-part board of which roughly a fifth are active devices, a shortage can make almost every IC and transistor unobtainable at once, forcing a choice between redesigning around what is available—which may take longer than the shortage lasts—and buying through the grey market at whatever price is asked.[601] Every forced substitution carries a design cost beyond the part price: the board must be reworked, sometimes repeatedly, and each change reduces confidence in the result, which must be bought back with additional testing or accepted as risk.[573]
Supplier behaviour is itself a selection criterion. Longevity assurances are only as durable as the corporate structure behind them: a letter promising ten years of availability is worth nothing if the company is subsequently broken into separate businesses.[118] A supplier can also fail on logistics rather than engineering—one manufacturer released new devices at a rate of hundreds per year, faster than customers could track, then could not deliver them, with available capacity going to the highest-volume customers and smaller designers who had specified the parts left unable to obtain them.[44] Manufacturer support quality is regional in a way that affects where hardware development happens: in some regions a device question unresolvable from the data sheet, such as whether a pin has an internal pull-up, is answered by a local representative within a day or two, a level of response not available everywhere.[279]
Design and fabrication
Pin count constrains how much of a system fits on one die and thereby dictates architecture: a fixed number of channels per package forces a split into a control device and separate function devices, and two teams working independently on the same problem have arrived at near-identical partitioning because the constraints funnel the choices.[504] Programmable logic is designed by defining a tile and replicating it in two dimensions to whatever size the product requires, so the difficult work is the tile and the chips that follow are a multiplication of it.[103] Process advances arrive as combinations of dimension and voltage rather than dimension alone: moving from a 22-nanometre process to a 16-nanometre fin-based one came with a core supply voltage around 0.55 volts, which changes the surrounding design as much as the density does.[246]
A mask error has no field remedy, since there is no equivalent of a wire modification on a die—“There ain’t no bodge in a chip”—which separates a silicon mistake from a board mistake in consequence rather than in kind.[455] Making a usable integrated circuit outside an industrial setting is separated from making a single device by many orders of magnitude of complexity: producing one transistor or one light-emitting diode is achievable, but producing something with logic levels, packaging, leads and pins that can be designed into a circuit is the actual problem, and the market’s abundance of cheap qualified parts removes most reasons to try.[75]
What lowers the barrier to custom silicon is documentation and access rather than equipment. A process development kit (PDK) is the fundamental information needed to create an integrated circuit on a given process, and an open PDK combined with shared fabrication runs allows designs to be made without a foundry relationship.[703] An open silicon programme can permit commercial use provided the design itself is open, on the reasoning that a healthy ecosystem needs companies contributing to it and building the unglamorous parts, in the same way commercial contributors sustain widely used open-source software.[501] At the industry level, the separation between design and manufacture is structural: a major supplier such as Arm sells intellectual property rather than devices, a distinction that surprises those who assume every named chip company operates fabrication.[625]
Engineering practice
The working assumption during debugging is that the IC is behaving correctly, and the chip is discounted last; the assumption is right most of the time but not always—a fault initially attributed to capacitors and then to the supply converter turned out to be a failed chip presenting as two ohms across the power rail.[4] Silicon defects are universal across manufacturers rather than characteristic of any one of them, so choosing a vendor to avoid them is futile; even a modest eight-bit device contains an enormous number of transistors and circuits, and its maker has already moved on to the next design.[482] A characteristic failure mode is back-powering: a signal driven into an IC whose supply is off passes current through the protection diodes present on every input and output, partially energising the internal rail, so a device that should be unpowered is partly alive and behaves unpredictably.[527]
Simulation of a design imported from schematic capture stalls on the active devices: passive components map onto generic models automatically, while each integrated circuit must exist in the simulator’s library or have a model chosen by hand.[148] Documentation practice requires that a schematic name the device, not merely carry a reference designator, since forcing a reader to cross-reference a bill of materials defeats the purpose of the drawing.[80] Transient protection is applied in two places rather than one—where cables enter the product, and again at the sensitive devices themselves—on the reasoning that the entry protection will not catch everything.[165] Thermal behaviour is instrumented by placing small resistance temperature detectors directly onto the packages of interest, resolving heat sources, sinks and the thermal path through the assembly rather than giving a single board temperature.[425] Prototyping with milled boards is used specifically to characterise unfamiliar devices: a quick board exists to learn how an IC behaves and to develop a working driver, after which a conventional board is ordered.[245]
Devices manufactured decades ago remain functional after long storage: parts pulled from foam after twenty or thirty years work when fitted, consistently enough to be treated as the expectation rather than good luck.[246] Products nonetheless fail more often at a connection than at a semiconductor—at a connector or a wire—and connector selection and maintenance receive correspondingly less attention than they warrant because integrated circuits attract the interest.[708]
The applications engineering role at a semiconductor company spans the whole life of a part: helping define its features and pin-out, writing the data sheet, supporting customers and field engineers through early design-ins, validating and debugging the silicon, and releasing it to production before starting the next one.[566]
Analysis and security
Integrated circuit security research borrows its equipment from failure analysis laboratories attached to fabrication plants, using the same instruments the manufacturer uses to inspect a die in order to manipulate or read the chip in ways its designers did not anticipate, after opening the package to expose the bare die.[303] Reading state out of a modern chip with a laser pulse and observing whether a bit flips is conceptually the same technique as reading magnetic core memory, except that core reads destroyed the stored value and required it to be written back, whereas the optical technique need not be destructive unless pulse energy is raised to the point of damage.[303] Recovering a circuit from a finished die is done by etching away the top and working downward through the metal layers, redrawing the connectivity as each is removed; the method is straightforward and the labour is substantial.[128]
The integrated circuit occupies a specific place in this landscape because circuit design and software engineering are the same activity of building abstractions: a physical phenomenon is distilled into a discrete component, components are assembled into an integrated circuit with a defined interface whose contents the user is not expected to know, and software repeats the process from machine language upward; security work consists largely of finding the boundary conditions where those abstractions fail, including at the hardware boundary itself.[725]