Optimizing Custom Single-Purpose Processors

Computer Organization & Embedded System — Hardware–Software Design Issues on Embedded System, NEC licence examination syllabus (Nepal Engineering Council).

Optimizing Custom Single-Purpose Processors

Once it works, make it smaller, faster, or cheaper.

🛠️ Three Optimization Targets

The FSMReduce the number of states (merge equivalent states — same idea as FSM minimization from digital logic), and choose efficient state encoding to reduce decoding logic.

The datapathShare one functional unit (like one adder) across multiple operations instead of duplicating hardware, and reuse registers where lifetimes don't overlap.

Combinational logicSimplify Boolean expressions (same techniques as Boolean algebra/K-maps) to reduce the number of gates needed.

💡 Common question: "Given two operations that both need an adder but never run in the same clock cycle, how can you optimize the datapath?" → Share a single adder between them using a multiplexer to select its inputs.

total cost = NRE + (unit cost × volume) With v = 10000, custom = 520000, gpp = 170000.

The two lines cross near 37,000 units. Below that the GPP wins on cost; above it the custom design does.

The calculation that decides custom versus general-purpose

Every trade-off in this section — flexibility, performance, power, NRE — comes down to one piece of arithmetic that the comparison tables leave implicit.

NRE (non-recurring engineering) cost is paid once no matter how many units ship: design, verification, and for a custom chip the photomasks. Unit cost is paid per device. So:

total cost = NRE + (unit cost × volume) Custom design: NRE = $500,000, unit = $2 GPP: NRE = $20,000, unit = $15 Break-even where the two are equal: 500,000 + 2v = 20,000 + 15v 480,000 = 13v v = 36,923 units

Below roughly 37,000 units the general-purpose processor is cheaper overall, however inefficient it is per unit — the custom design's NRE simply never gets amortised. Above it, the $13 saved on every unit overwhelms the up-front cost and keeps widening the gap.

Checking a few volumes: 1,000 units: custom $502,000 vs GPP $35,000 → GPP 10,000 units: custom $520,000 vs GPP $170,000 → GPP 50,000 units: custom $600,000 vs GPP $770,000 → custom 100,000 units: custom $700,000 vs GPP $1,520,000 → custom

💡 This is why a consumer product shipping millions of units gets a custom chip while industrial equipment shipping hundreds uses an off-the-shelf microcontroller — the same engineering decision resolves in opposite directions purely on volume. Quote the break-even formula in a long answer; it is the reasoning the table is summarising.

What the cost model leaves out

Cost is not the only axis, and a full answer says so.

⏱️ The other pressures

Time to marketA custom chip takes months to fabricate. Missing a launch window can cost more than the entire NRE saved.

FlexibilityA bug in a GPP's software is a firmware update. A bug in fabricated silicon is a new mask set and another full NRE.

PowerCustom hardware does only what is needed, so it can be far more efficient — decisive in battery-powered products regardless of volume.

PerformanceA custom datapath does in one cycle what a GPP needs many instructions for.

ASIPs sit deliberately in between: a programmable core with instructions added for one application domain. They carry lower NRE than a full custom design while beating a general-purpose processor on the workload they target — which is why the four-way comparison has four entries rather than two.

Syllabus points

Optimizing the FSM
Optimizing the datapath
Optimizing combinational logic

Create a free account to tick topics off, take notes as you read, watch the video lessons and get a day-by-day study plan built around your exam date.