VLSI Interview Questions: MIPS/mm2, MIPS/Watt and Watts/cm3

This classic VLSI and computer architecture interview question asks whether three marketing metrics, MIPS/mm², MIPS/Watt and Watts/cm³, are sensible for customers choosing a system. The short answer: MIPS/Watt is reasonable (bigger is better, most relevant to cell phones), Watts/cm³ is reasonable (smaller is better, most relevant to compute servers), and MIPS/mm² is mainly a designer's metric rather than a customer's one. The interviewer is really testing whether you can tie each metric to the cost or constraint a buyer actually cares about.

Key takeaways

• MIPS/mm² (performance per unit of silicon area) matters to chip designers because area drives manufacturing cost, but customers rarely buy on it directly.
• MIPS/Watt (performance per unit of power) is a real customer metric: bigger is better, and battery-powered phones care most.
• Watts/cm³ (power, and therefore heat, per unit of volume) is a real customer metric: smaller is better, and densely packed servers care most.
• Always justify the answer by naming the constraint: cost, battery life or cooling.
• MIPS itself is a weak performance measure across different instruction sets; strong candidates say so.

The question

One day you are strolling the hallways in search of inspiration, when you bump into a person from the marketing department. The marketing department has been out surfing the web and has noticed that companies are advertising the MIPS/mm², MIPS/Watt and Watts/cm³ of their products. This wide variety of different metrics has confused them.

Explain whether each metric is a reasonable metric for customers to use when choosing a system. If the metric is reasonable, say whether "bigger is better" (for example, 500 MIPS/mm² is better than 20 MIPS/mm²) or "smaller is better" (20 MIPS/mm² is better than 500 MIPS/mm²), and which one type of product (cell phone, desktop computer or compute server) the metric is most relevant to. Justify your answers.

Answer summary

Metric	What it measures	Reasonable for customers?	Better direction	Most relevant product
MIPS/mm2	Performance per unit of die area	No (a design metric)	Bigger, for the designer	Weak case for cell phones
MIPS/Watt	Performance per unit of power	Yes	Bigger is better	Cell phone
Watts/cm3	Power (heat) per unit of volume	Yes	Smaller is better	Compute server

MIPS/mm²: reasonable? No

MIPS/mm² measures performance per unit of area. Although we often make performance and area trade-offs on a particular chip, the reasons for being concerned about area are based on the manufacturing cost of the chip. Area by itself is generally not a primary concern for the person buying the system; they see the cost, not the square millimetres.

There is one semi-plausible argument for this metric. When looking at chips for cell phones, we could trade off small size and area, because the size of a cell phone is an important concern for consumers. Even then, the phone's size is set far more by the battery, display and board than by the processor die, so the argument is weak.

Why designers still care

Inside a chip company, performance per area is a useful figure of merit. A smaller die means more dies per wafer and, at a given defect density, a higher yield, so cost per good die falls faster than area alone suggests. As an illustrative example, halving the area of a block does not just halve its silicon cost; it also lowers the chance that a random defect lands on it. That is a supplier's concern, which is why the customer-facing answer is still "no".

MIPS/Watt: reasonable? Yes

This metric measures performance per unit of power consumption. In a market segment that is concerned about both performance and power, it is of direct concern.

Which is better? Bigger is better. Increasing performance and decreasing power consumption are both good things.

Most relevant product? Cell phone. Of the three candidate products, cell phones are the most sensitive to power consumption, because power drains the battery and also has to leave the device without a fan. Phone processors were also the most likely of the three to be marketed with MIPS as the measure of performance. Desktop PCs and compute servers generally use SPEC benchmarks for performance.

A worked example

Illustrative numbers only: chip A delivers 2,000 MIPS at 1 W (2,000 MIPS/W), and chip B delivers 3,000 MIPS at 2 W (1,500 MIPS/W). Chip B is faster, but for the same task chip A uses less energy. Since energy per instruction is simply the inverse of MIPS/W, chip A needs 0.5 nJ per instruction and chip B needs about 0.67 nJ. For a phone that spends a fixed amount of work per day, chip A gives longer battery life.

A good follow-up point: MIPS/W is the same as instructions per joule, which is why energy-efficiency metrics such as performance per watt are used for servers too. The difference is that a server is limited by electricity and cooling cost, whereas a phone is limited by battery and skin temperature.

Watts/cm³: reasonable? Yes

This metric measures power consumption, or heat generation, per unit of volume. When heat generation is a concern, the primary concern is how much heat is generated per volume of space, because that heat must be carried away. There are three broad categories of cooling: no active cooling needed, air cooled (for example, fans) and liquid cooled (for example, water or other coolant loops, with liquid nitrogen reserved for extreme overclocking experiments).

Which is better? Smaller is better. The less heat that is generated per volume of space, the less need there is for active cooling, and the more comfortable the temperature is.

Most relevant product? Server. Cell phones generally don't generate enough total heat for this metric to be the main concern. Desktop PCs generally are not packed into a room so tightly that the volume of the PC matters. Compute servers tend to be packed into rooms and onto racks as densely as possible, so their volume and heat dissipation are of concern.

Related metric: Watts/cm²

Engineers also quote power density per unit of area, in W/cm², at the die or package level. That figure drives the choice of heat sink and thermal interface material. Watts per volume is the system and data centre view; watts per area is the chip and package view. Mentioning both shows you understand where the heat goes.

A caveat about MIPS itself

MIPS (millions of instructions per second) only compares fairly between processors that run the same instruction set and the same program. A complex instruction may do the work of several simple ones, so a higher MIPS figure does not always mean more useful work. This is why desktop and server buyers rely on application benchmarks such as the SPEC CPU suites, and why a careful answer treats any MIPS-based metric with some suspicion.

How interviewers probe this question

• "Who is the customer?" They want you to separate the chip designer's view (area, yield, cost) from the system buyer's view (battery life, cooling, running cost).
• "Why not MIPS/Watt for servers?" A good answer: servers do care about performance per watt, but they measure performance with benchmarks rather than MIPS, and rack density adds the volume concern.
• "Is energy per operation the same as MIPS/W?" Yes, it is the reciprocal, which is often the clearer way to reason about battery life.
• "What limits power in a phone?" Battery capacity and the temperature of the case in the user's hand, since there is no fan.

Common mistakes

• Saying "bigger is better" for every metric without thinking about whether the quantity in the numerator is good or bad.
• Treating die area as a customer concern rather than a cost driver.
• Picking the desktop PC as the most relevant product for any metric; it is rarely the most constrained of the three.
• Forgetting to justify the answer, which is what the question actually marks.

FAQ

Is MIPS/mm² ever useful?

Yes, to chip architects comparing core designs on the same process, because area sets cost. It is just not a good basis for a customer's buying decision.

Why is MIPS/Watt most relevant to phones rather than servers?

Both care about efficiency, but the phone runs from a battery with no active cooling, so power is the hardest limit it faces. Servers usually express performance with benchmarks rather than MIPS.

Why does a server care about heat per volume?

Data centres pack servers densely into racks, and the cooling system must remove all the heat produced in that space. Higher W/cm³ means more expensive cooling or fewer servers per rack.

What is the difference between W/cm² and W/cm³?

W/cm² is power density across a die or package surface and drives heat sink design. W/cm³ is heat per unit of enclosure volume and drives rack and room cooling.