| Episodes | 61 |
| Mentions | 73 |
| Cited here | 49 |
| First — last | #19 — #706 |
| Top guests | Michael Ossmann, Jonathan Ellis, Avidan Ross |
| Related | altium · intellectual property · kickstarter · oscilloscope · consulting |
Research and development, commonly abbreviated R&D, is the organised effort by which companies, laboratories and individuals produce new technical knowledge and convert it into products. The two halves are not equivalent: research cannot be held to a predicted payoff, since demanding a specific return from a research programme converts it into development.[200] Development dominates in practice — a commonly cited cost ratio holds that development costs at least ten times as much as research and commercialization ten times as much as development, leaving research at roughly 1 percent of the total effort of bringing a technology to market.[104] Most engineering employment is accordingly development work aimed at producing an answer and getting a design out the door, because a company will not hire enough engineers to keep someone available to invent the next product.[385]
Research and development distinguished
Startups perform development rather than foundational research, because research is expensive, fails often, and requires reserves a young company does not have.[231] Assembling existing components into a difficult product is likewise development rather than fundamental research; firms that advance the underlying sensor technology occupy a different category from companies packaging those sensors.[225] Genuine research legitimately involves isolating a team to try many approaches and accepting failure as part of the process, since the outcome cannot be specified in advance.[205]
The distinction has direct consequences in contract engineering. Projects that are genuinely new work from scratch cannot be priced as a fixed bid, because the work sits closer to science than engineering and its endpoint is unknown when the contract is signed.[409] Consulting on unsolved problems means taking a client’s money for work that may not succeed, which is the practical boundary between research and development in contract work.[549] Doctoral programmes present the same choice in career form: a student may leave with a master’s degree after several years, so the decision to continue turns on how much the candidate values research over development-type engineering.[164]
The categories can also be mixed destructively. Substituting exploration for design — having a domain expert write and play with code until behaviour seems right and then freezing it — is research being run inside a development programme; in a regulated medical device it produced unending defects, an FDA investigation and millions of dollars in cost.[489]
Funding
The relationship between R&D spending and output is not monotonic: a small amount of money produces nothing because the infrastructure is missing, and beyond a point additional money stops correlating with output per dollar.[74] The floor is nonetheless absolute, since zero R&D funding yields zero R&D results, so arguments about diminishing returns apply only above some non-zero level of investment.[74] The finding that more money does not mean more output is routinely used out of context to justify budget cuts, ignoring that marginal spending buys concrete capability such as a better oscilloscope, extra parts, or another laboratory technician.[74]
Spending expressed as a percentage of revenue is a misleading comparison between companies: a firm spending 2 percent of a very large revenue can outspend competitors whose combined budgets are smaller, and a low percentage may reflect development offloaded onto suppliers.[74] A company that announces an increase in dollars also leaves open whether the money goes to equipment or capital expenditure, whereas a stated headcount increase is a far more specific commitment to engineering capacity.[28]
Several structures fund the work indirectly. High-end instruments carry the margins that pay for an instrument maker’s R&D, though retreating from the low and mid range costs the vendor its presence on working benches and the customers who would later buy up the range.[164] Corporate funding of university research is cheap relative to internal headcount, with about 150,000 a year covering roughly one full-time employee plus parts, and companies expecting breakthrough ideas rather than quarterly key performance indicators in return.<sup><a href="#ref-283" title="Ep 283: An Interview with Jonathan Ellis (January 20, 2016)">[283]</a></sup> Government funds it directly as well: DARPA, the Defense Advanced Research Projects Agency, is part of the United States Department of Defense and funds research and development that is primarily military-focused but sometimes broad, and its predecessor ARPA created the internet.<sup><a href="#ref-177" title="Ep 177: Discussing Innovation and the Future with Mike Ossmann - Fiesty Festivus Futurology">[177]</a></sup> Grant funding restricts how recipient time is used, prohibiting salaried hours from being charged to sales and marketing work.<sup><a href="#ref-330" title="Ep 330: An Interview with Zach Fredin (January 4, 2017)">[330]</a></sup> Some large manufacturers substitute prize contests for internal effort, as with a competition paying up to 10,000 for integrating a flexible screen, where entrants should read the terms to see what rights they sign over.[160]
Spending on research into new markets is partly defensive, since a company that does not fund the search risks a competitor reaching the market first, and a failed programme can at least be written off as a market that did not exist.[79] Continuously chasing new markets is not the only viable strategy, however; a company that owns a market and keeps innovating inside that space has a defensible position.[134]
Organisational settings
Corporate research laboratories vary in charter. One can be chartered explicitly not to feed the product pipeline, instead being asked to look three to five years ahead at how emerging social, cultural and technological movements combine, which permits output ranging from CNC-heavy mechanical installations to machine learning work.[204] Large research organisations such as MIT Lincoln Laboratory, with roughly 3,000 staff, assign a second team of senior engineers to review the team executing a major programme, producing a culture of formal and informal design reviews that can become contentious.[119] Amazon’s Lab126 developed the Kindle and later the Fire Phone, and the failure of the phone changed the character of what the group was permitted to work on.[231]
No present-day equivalent of Xerox PARC or Bell Labs exists as a single broad-front research institution; long-horizon work is instead split across many small groups working on specific areas such as artificial intelligence and quantum computing, partly because modern problems span more disciplines.[361] A large passive component manufacturer illustrates a distributed alternative, separating central R&D at a single site — in AVX’s case Greenville, South Carolina — from applied R&D and process R&D spread across its worldwide plants.[596]
Organisational size determines how the work feels. Very small companies and startups impose almost no procedural overhead on development, which is the principal structural contrast with large defence contractors.[588] Thin staffing at small companies also produces unusually broad individual scope: at a synthetic biology startup of roughly two dozen people, a single engineer wrote the software for the hardware and also spun the PCBAs for factory instruments.[611] A salaried engineer at a large organisation can pursue longer-term work that probably will not succeed, whereas someone self-employed or close to manufacturing is constrained to work that makes money today or protects revenue already coming in.[227]
R&D capability is unevenly distributed across industries. Established industries outside technology often carry expensive operational problems and substantial revenue while having no internal R&D department, so they look to outside technology companies to supply innovation.[327] Industrial customers frequently outsource development to consultancies from rough sketches.[385] Unconventional arrangements exist as well: a media channel can serve as the development arm of a product business, with public build videos acting as the R&D function that feeds the manufacturing and retail side.[592]
Where investment rather than employment supplies the funding, a hardware-focused venture fund may place about a third of its capital in consumer and deliberately co-invest with social media, industrial design or branding investors to cover the skills it lacks, in contrast to funds chasing a single outsized win to offset a hundred failures.[327]
Semiconductor and component development
Analog IP providers can spend roughly half their engineering hours on internal test-chip development rather than customer deliverables, because staying at the leading edge depends on running silicon of their own.[706] Foundries support such work by offering discounted development area on multi-project wafers, which lets small teams get test designs fabricated without funding a full mask set.[706] Semiconductor research can be done on one-centimetre wafers grown in a laboratory, but a real manufacturing process requires four- or six-inch wafers, so scaling the substrate is a distinct step between research and production.[71]
Once a device exists, its economics change character. Custom ASICs make advanced low-cost instruments possible because the R&D funding is a one-time cost: after the chip exists each unit costs only a few dollars to produce, which smaller competitors without that capability cannot match.[30] Agilent moved its low-end oscilloscopes from rebadged Rigol units to in-house design and manufacture, reversing a decade-long pattern of outsourcing by large American instrument makers.[30]
Semiconductor suppliers will go beyond their catalogue of standard building blocks for a customer if the volume justifies it, but the engineering hours they commit are billed back to the customer.[129] When a semiconductor company under financial pressure strips R&D, the cut tends to fall on the legacy line rather than the strategic one, so 8-bit microcontroller development is sacrificed to keep 32-bit families funded, stranding usable existing IP such as PowerPC and ColdFire.[116]
Component development has its own history. Capacitor work around forty years ago could be run on the production line itself, with single-layer parts made by taking MLCC scrap from the edges of wafer starts, grinding it down by hand and metallizing it, an improvisation no longer possible on modern lines.[596] Present-day equivalents are considerably more structured: hermetic tantalum polymer capacitors are built by welding tantalum polymer pellets into a ceramic package processed like a standard hybrid microcircuit and sealing it with a non-moisture-bearing internal atmosphere.[596]
In design software, EDA tool vendors run on small teams with limited engineering bandwidth, so effort concentrates on the core tool and component data becomes a secondary concern, leaving room for specialist suppliers of footprints and symbols.[531] Incumbent tools are difficult to displace because each year’s spending of hundreds of millions on development compounds on roughly twenty years of accumulated prior IP, alongside vendor relationships that feed the incumbent directly.[645]
From laboratory to product
Semiconductor performance records announced by large research laboratories do not correspond to purchasable parts; exotic processes such as gallium arsenide only reach ordinary products years later, once the process has spread and cheapened.[17] A genuinely new semiconductor technology typically takes ten or more years from research demonstration to being stocked in one-off quantities by distributors such as Mouser or Digi-Key.[296] Laboratory demonstrations should also be assessed against the density and volume the target market needs: a memory technology demonstrated at kilobyte densities has not been shown at the megabit densities a commercial product requires.[104]
Research-weighted work requires many iterations through different ideas; a concept prototyped in a day demonstrates only a limited case, and expanding it to work in every case can take on the order of twenty iterations.[141] Work that pushes electronics beyond straightforward component integration runs on three-to-four-year product cycles rather than the six-month cycles of consumer electronics, and depends on management patient enough to accept that the result has not been done before.[305]
Much consumer hardware sits at the other extreme, being not R&D-centric but a repackaging of silicon vendors’ reference designs, with the vendor absorbing the development cost; many consumer routers are the networking chip’s reference circuit rebuilt.[363] Very low part cost generates experimentation independently of product polish, since modules such as Espressif’s and LoRaWAN hardware are not especially refined but the near-absence of a price barrier puts them in far more hands.[403] A product requiring little development effort can still succeed rapidly through creative marketing, making the marketing capability rather than the technology the competitive advantage.[394]
Development practice
One method for generating project ideas is to keep a wide inflow of unrelated work and repeatedly ask which pieces are close to which others and whether they can be combined; most such combinations are bad, and only occasionally does one become developable.[232] Starting from a technology and hunting for an application instead produces a solution looking for a problem; the more reliable direction is to find problems people already have and check whether current work solves them.[232] Engineers can also steer their careers by seeking work that justifies expensive components, since the interesting problems live there, whereas shaving a MOSFET cost from thirteen cents to twelve and a half cents is a different and narrower class of challenge.[232]
Breaking a project into sub-assemblies and finishing each one individually before integration protects progress against interruptions, since a project worked on outside full-time hours will be interrupted and a half-finished whole is easier to abandon than a finished sub-assembly.[550] Releasing a project so that others can build it is a separate project in its own right, requiring instructions, documentation, comment handling and ongoing support, and should be scoped as such rather than treated as the tail end of the build.[550]
Design constraints increasingly come from outside the technical requirements. Major consumer electronics manufacturers specify recycled content and disposal hazard as design constraints, advertising products built with around 30 percent recycled plastics.[24] Restraint in scope can itself be the decisive design choice: restricting the Kindle to reading books is what allows battery life measured in months.[231] Once the novel phase of a hardware product is finished, the remaining work is refinement and cost reduction, which offers neither the technical challenge nor, for an already-profitable designer, the financial motivation of the original development.[194]
Technology choices are sometimes made for workforce reasons rather than technical ones. Ladder logic persists in industrial control because it lets a technician with about three weeks of training programme a multimillion-dollar machine, and that interchangeability of personnel also explains vendor concentration around Allen-Bradley.[385]
Licensing third-party IP is low-risk where the block is not the basis of differentiation, such as video compression in a specialised product whose advantage lies elsewhere, and writing the same functionality from scratch is a deliberate R&D cost taken against a long-term payoff.[536] The point at which a team should absorb the cost of replacing bought-in IP with its own implementation is around the second or third revision, once the product has proven itself.[536]
Contract research and development
A consultant should enter a fixed-bid contract only when both a deliverables list and acceptance criteria exist, meaning the specific tests and outcomes that define completion; with both in hand, estimates can land within 5 to 10 percent.[409] Where requirements and acceptance criteria do not yet exist, an hourly agreement with a check-in every ten hours lets the client approve or stop the work in bounded increments while the consultant buys the exploratory time needed to write a specification, after which the engagement can convert to fixed bid.[409] A practice structured this way can settle at roughly three quarters fixed-bid work and one quarter hourly.[409]
Contract R&D is exposed to its clients’ budget cycles. In a downturn, large companies cut research and development budgets first, and the cancellations propagate immediately to the 20- to 30-person design and development consultancies that serve them, which is how a functioning service business can collapse within months.[402]
Returns and outputs
Patents function as the return on R&D expenditure: a company that spends several million dollars on engineers and prototypes ends with a saleable and defensible asset it can license, sell, or use to protect the revenue that justifies the investment.[420] The same instrument supports a separate business: in legal terminology a non-practicing entity is a party that buys patents with no intention of using them, and because patent litigation is very expensive to defend, nuisance suits can extract settlements smaller than the cost of fighting them.[420]
Value can also leave the organisation that paid for it. Ethernet was invented at Xerox PARC, but the technology was commercialized only after members of that team left to found 3Com.[361] The same laboratory built the Xerox Alto as a ten-year lookahead rather than a product, at an effective cost of roughly $100,000 per machine, an approach vindicated when the Macintosh appeared in 1984, about eleven years later, closely matching what the team had projected.[361]
Open exploratory hardware projects, such as successive experiments with PCB coil propulsion, function as research whose payoff depends on someone later finding an application for the technique.[403]
Constraints and failure modes
Administrative controls calibrated for production can paralyse laboratory work, as when a production facility required a thorough written justification before a second $100 multimeter could be bought.[200]
Supply conditions suppress exploratory work directly: when lead times force immediate purchasing decisions, there is no time to do the research and then build a prototype, so planned investigations are skipped.[552] Under such constraints small hardware companies must buy a hundred or two hundred parts, potentially a year’s supply, for designs they are only fairly confident about and whose production may be far off, which strains businesses that normally fund each production run from the proceeds of the last.[552]
Regulatory cycles impose a similar tax on iteration. A certified medical device whose release cycle requires three months of qualification per change makes iterative code exploration ruinously expensive, since even a one-line change carries the full three-month cycle.[489]
The absence of results is itself diagnostic. A development programme that has run five years without a single measurement of the two functions its product must perform, such as solar output and stopping a vehicle, has not been testing at all.[205] Copied material carries its own hazard: vendor example software is written for ease of evaluation and carries explicit warnings against production use with security features disabled, so shipping copied example code, such as a bundled web server, is a recognised source of insecure devices.[363]