These questions were asked in a Hughes Networks interview for VLSI design roles. They cover the core of front-end and back-end digital design: setup and hold, scan and DFT, equivalence checking, CMOS delay, power and process variation. Each question below has a short answer first, and the deep dive works through the numbers.
The questions
- What is setup/hold? What are setup and hold time impacts on timing? How will you fix setup and hold violations?
- Explain the function of a muxed FF (multiplexed flip-flop) / scan FF (scan flip-flop).
- What is tested in DFT (design for testability)?
- In equivalence checking, how do you handle the scanen signal?
- In terms of CMOS (complementary metal oxide semiconductor), explain the physical parameters that affect propagation delay.
- What are the power dissipation components? How do you reduce them?
Answer:
Short Circuit Power
Leakage Power Trends
Dynamic Power
Low Power Design Techniques
- How is delay affected by PVT (process, voltage, temperature)?
Answer: Process-Voltage-Temperature (PVT) Variations and Static Timing Analysis (STA)
- Why is the power signal routed in the top metal layers?
Source: vlsifaq.blogspot.com
Short answers
Setup and hold
Setup time is how long data must be stable before the capturing clock edge; hold time is how long it must stay stable after it. A setup violation means the path is too slow for the clock period. Fix it by reducing logic depth, upsizing cells, moving logic across registers, using faster (lower-threshold) cells, adjusting clock skew in your favor or, as a last resort, lowering the frequency. A hold violation means data races through too fast and corrupts the value being captured. Fix it by adding delay (buffers or slower cells) on the short path or reducing the skew that causes it. Hold does not depend on clock frequency, so slowing the clock never fixes it.
Muxed scan flip-flop
A scan flip-flop is a normal D flip-flop with a 2:1 multiplexer on its input. When scan enable (SE) is 0, the flop captures functional data D. When SE is 1, it captures scan input SI, which is wired to the previous flop's output. All the scan flops then form a shift register, so a tester can shift a pattern in, run one functional clock, and shift the result out.
What DFT tests
Manufacturing defects, not design bugs. Scan with automatic test pattern generation (ATPG) targets stuck-at faults (a node stuck at 0 or 1) and transition (at-speed) faults. Memories are tested with built-in self-test (MBIST). Boundary scan (JTAG, IEEE 1149.1) tests the chip's pins and board connections. Some flows also use IDDQ (quiescent current) testing and path delay tests.
Scan enable in equivalence checking
Scan insertion changes the netlist, so the scan-inserted netlist is not equivalent to the RTL unless scan mode is disabled. Constrain scan enable (and any test mode pins) to their functional value, normally 0, in the implementation, so the scan multiplexers always select the functional path. Then the comparison checks only functional logic.
Physical parameters that affect CMOS delay
Gate delay is roughly the load capacitance times the voltage swing divided by the drive current. The current rises with transistor width W, falls with channel length L, rises with oxide capacitance (thinner oxide), carrier mobility and supply voltage, and falls with threshold voltage. Load capacitance comes from the fan-out gate capacitance, the wire and the driver's own junctions. Temperature changes both mobility and threshold voltage.
Power dissipation components
Dynamic switching power (charging and discharging load capacitance), short-circuit power (both transistors briefly on during an input transition), and leakage (subthreshold, gate and junction leakage). Reduce them with lower supply voltage, clock gating, lower switching activity, sharp input transitions, high-threshold cells on non-critical paths and power gating of idle blocks. The linked articles above cover each in detail.
PVT and delay
Slow process corners, lower supply voltage and, for most older nodes, higher temperature all make gates slower. At low supply voltages in advanced nodes the temperature effect can reverse (temperature inversion), so the slowest corner may be the cold one. STA is run at several corners to cover the worst cases for setup and for hold.
Why power goes on the top metal layers
The top layers are the thickest and widest, so they have the lowest resistance. That keeps IR drop and electromigration within limits for the large currents a power grid carries. They are also closest to the package bumps or bond pads, and putting the wide power straps up there leaves the thin lower layers free for signal routing.
Deep dive: the numbers and details
This part holds worked numbers for the questions above. Skip it if the overview is all you need. All values are illustrative.
Setup and hold slack
Setup slack = (T + T_skew) - (T_clk_to_q + T_logic_max + T_setup)
Hold slack = (T_clk_to_q + T_logic_min) - (T_hold + T_skew)
T_skew is the capture clock arrival minus the launch clock arrival. With T = 2.0 ns, clock-to-Q 0.15 ns, worst logic 1.70 ns, setup 0.10 ns and skew +0.05 ns: setup slack = 2.05 - 1.95 = +0.10 ns. If the shortest path through the same flops has 0.02 ns of logic and hold is 0.08 ns: hold slack = 0.17 - 0.13 = +0.04 ns. Increasing the skew to fix setup would eat directly into that hold margin.
Delay against supply voltage
A simple model (the alpha-power law) gives delay proportional to V / (V - Vth)^alpha, with alpha between 1 and 2 for short-channel devices. With Vth = 0.35 V and alpha = 1.3:
| Supply | V / (V - Vth)^1.3 | Delay relative to 1.0 V |
|---|---|---|
| 1.0 V | 1.0 / 0.65^1.3 = 1.75 | 1.00 |
| 0.9 V | 0.9 / 0.55^1.3 = 1.96 | 1.12 |
| 0.8 V | 0.8 / 0.45^1.3 = 2.25 | 1.29 |
A 10 percent supply droop costs about 12 percent in delay here, which is why STA uses the low-voltage corner for setup and why IR drop on the power grid matters.
Power components in numbers
P_dynamic = alpha x C x V^2 x f
P_leakage = I_leak x V
A block with activity 0.1, 1 nF of switched capacitance, 0.9 V and 800 MHz has dynamic power 0.1 x 10^-9 x 0.81 x 8 x 10^8 = about 65 mW. If it leaks 5 mA at 0.9 V, leakage adds 4.5 mW. Clock gating that halves the activity saves about 32 mW; power gating during idle removes most of the 4.5 mW.
IR drop on the power grid
Wire resistance is R = R_sheet x length / width. Suppose a thick top layer has a sheet resistance of 0.02 ohm per square and a thin lower layer 0.1 ohm per square. A 1,000 um long, 10 um wide strap is 100 squares: 2 ohms on the top layer against 10 ohms on the lower one. Carrying 10 mA, the drop is 20 mV against 100 mV. On a 0.9 V supply, 100 mV is over 11 percent, enough to slow gates by the amount shown in the table above.
Scan shift time
With 100,000 scan flops split into 100 chains of 1,000 flops and a 50 MHz shift clock, each pattern takes about 1,000 shift cycles, 20 microseconds. For 5,000 patterns that is about 0.1 seconds of tester time per chip. More chains (or scan compression) cut this proportionally, which is why chain count is a cost decision.
-------------------------------PROMOTION-------------------------------
Advertise on this Blog and see yourself grow!
For more information contact us at onenanometer[at]gmail.com


Your comments will be moderated before it appears here.