These Hynix Semiconductor interview questions, first published here in December 2008, focus on backend and timing skills: power optimization across the physical design flow, timing fixes before and after layout, process technology challenges, clock dividers, multicycle and false paths, hold violations, RTL coding for good backend results and timing diagrams. Hynix joined the SK Group and was renamed SK hynix in March 2012, as reported by bit-tech on March 27, 2012, and the questions remain standard for memory and logic design roles. Model answers follow each question.
Key takeaways
- Power answers should go stage by stage: architecture and RTL, synthesis, placement, CTS and routing, and always mention the power, performance and area (PPA) trade-off.
- Pre-layout timing fixes are structural; post-layout fixes are local (sizing, buffering, Vt swaps, useful skew).
- Odd and fractional clock dividers with 50% duty cycle need both clock edges; verified Verilog for divide by 2, 3 and 1.5 is below.
- Multicycle and false paths are constraint changes, and each one must be justified by the design, not by the timing report.
- Hold is checked on the fastest path and does not depend on the clock period.
What the interview covers, in plain words
The eight questions below test whether you understand how a chip goes from RTL to a working layout, and where timing and power are won or lost along the way.
- Power: power is saved at every stage, from switching off unused blocks in the architecture to choosing slower, less leaky cells where timing allows. The best answers go stage by stage and admit the cost in speed or area.
- Timing before and after layout: before layout you fix timing by changing the structure of the logic; after layout you fix it locally by resizing cells, adding buffers or moving cells.
- Process challenges: at smaller nodes, wires get slower relative to gates, variation grows, and leakage, noise and supply drop all get harder to control.
- Clock dividers: dividing a clock by an odd number or by 1.5 with a clean duty cycle needs both edges of the input clock.
- Timing exceptions: multicycle and false paths tell the timing tool which paths do not need to meet a single-cycle check. They must come from the design intent, never from a failing report.
- Hold violations: a hold failure happens when new data arrives too soon after the clock edge. It depends on the shortest path and on clock skew, not on the clock period.
- RTL for good backend results: code that registers outputs, avoids long priority chains and uses clean clock gating is much easier to place, route and close.
- Timing parameters: setup, hold, recovery, removal and minimum pulse width all describe windows around a clock edge in which a signal must stay still.
If you only have a few minutes, the key takeaways above cover the core of each answer. The full model answers, code and tables follow.
Deep dive: the numbers and details
This part holds the full model answers, the Verilog for the dividers, the timing tables and the constraint commands. Skip it if the overview is all you need.
1. How do you optimize power at various stages in the physical design flow?
Power optimization is an important aspect of physical design, and there are several techniques that can be used to optimize power at various stages of the physical design flow. Here are some common techniques:
- Transistor sizing: adjusting the size of each gate or transistor for minimum power while still meeting timing.
- Voltage scaling: lower supply voltages use less power but go slower.
- Voltage islands: different blocks can be run at different voltages, saving power. Signals between islands need level shifters.
- Power gating: inserting design structures that turn off the supply voltage to a circuit during idle periods where the circuit is not in use. This needs isolation cells on outputs and, where state must be kept, retention registers.
- Clock gating: turning off (gating) the clock during periods where functions are idle to eliminate switching activity and the associated dynamic power consumption.
The practical application of power optimization techniques involves the analysis and trade-off of power, performance and area (PPA). Power optimizations often adversely impact performance and area; therefore, it is essential to analyze and optimize these trade-offs.
| Stage | Typical power techniques |
|---|---|
| Architecture and RTL | Power domains, clock gating enables, operand isolation, fewer transitions on wide buses |
| Synthesis | Automatic clock-gating insertion, multi-Vt libraries (high-Vt by default), area recovery |
| Floorplan and power plan | Voltage islands, power switch placement, level shifter and isolation placement |
| Placement | Placing high-activity nets short, downsizing cells off the critical path, Vt swaps |
| CTS | Fewer clock buffers, gating close to the source, controlled clock transition |
| Routing and sign-off | Shorter wires on high-activity nets, leakage recovery after timing closure |
The relation behind most of these is dynamic power P = α C V² f: reduce activity (α), capacitance (C), voltage (V) or frequency (f). Because voltage appears squared, voltage scaling usually gives the biggest gain, which is why voltage islands and DVFS matter so much.
2. What timing optimization strategies do you employ in pre-layout and post-layout stages?
Pre-layout (synthesis): correct and complete constraints first, then logic restructuring, retiming and pipelining, choosing faster architectures (for example a parallel-prefix adder instead of ripple carry), path grouping for critical paths, and multi-Vt optimization. Wire load models or physical-aware synthesis give an early view of wire delay.
Post-layout: gate sizing, buffer insertion and net splitting, swapping to low-Vt cells on critical paths, cloning high-fanout drivers, moving cells closer, layer promotion of long nets to thicker metals, useful skew in CTS, and hold fixing with delay cells. Late fixes are made as ECOs to avoid disturbing closed areas.
3. What are process technology challenges in physical design?
- Wire resistance rising faster than gate delay falls, so interconnect dominates timing.
- Larger on-chip variation, requiring OCV, AOCV or POCV style derates.
- Crosstalk and noise from tighter wire spacing.
- IR drop and electromigration as wires get thinner and current density rises.
- More complex design rules, multiple patterning and density requirements.
- Leakage power, which grows sharply at smaller geometries and higher temperature.
4. Design divide by 2, divide by 3 and divide by 1.5 counters. Draw timing diagrams.
Divide by 2 is a toggle flip-flop. Divide by 3 with a 50% duty cycle needs a pulse that lasts 1.5 input periods, which is made by ORing a positive-edge pulse with a copy delayed by half a period on the negative edge. Divide by 1.5 uses the same idea: one pulse from the positive edge and one from the negative edge, giving an output period of 1.5 input periods. A 50% duty cycle is impossible at divide by 1.5, because the output would need edges every 0.75 input periods, finer than the half-period resolution of the two edges; the design below gives a 2/3 duty cycle.
// Divide by 2: a toggle flip-flop
module div2 (input wire clk, input wire rst_n, output reg clk_out);
always @(posedge clk or negedge rst_n)
if (!rst_n) clk_out <= 1'b0;
else clk_out <= ~clk_out;
endmodule
// Divide by 3 with 50% duty cycle: a posedge pulse ORed with a negedge copy
module div3 (input wire clk, input wire rst_n, output wire clk_out);
reg [1:0] cnt;
reg p, n;
always @(posedge clk or negedge rst_n)
if (!rst_n) begin cnt <= 2'd0; p <= 1'b0; end
else begin
cnt <= (cnt == 2'd2) ? 2'd0 : cnt + 2'd1;
p <= (cnt == 2'd2); // high for one input period
end
always @(negedge clk or negedge rst_n)
if (!rst_n) n <= 1'b0;
else n <= p; // same pulse, half a period later
assign clk_out = p | n; // high for 1.5 periods out of 3
endmodule
// Divide by 1.5: output period is 1.5 input periods (duty 2/3)
module div1p5 (input wire clk, input wire rst_n, output wire clk_out);
reg [1:0] cnt;
reg a, b;
always @(posedge clk or negedge rst_n)
if (!rst_n) begin cnt <= 2'd0; a <= 1'b0; end
else begin
cnt <= (cnt == 2'd2) ? 2'd0 : cnt + 2'd1;
a <= (cnt == 2'd2); // high in the cycle where cnt is 0
end
always @(negedge clk or negedge rst_n)
if (!rst_n) b <= 1'b0;
else b <= (cnt == 2'd1); // high from mid cycle 1 to mid cycle 2
assign clk_out = a | b;
endmodule
Timing in units of the input period T (positive edges at 0, 1, 2 and negative edges at 0.5, 1.5, 2.5):
| Divider | Output high | Output low | Period | Duty cycle |
|---|---|---|---|---|
| Divide by 2 | 0 to 1 | 1 to 2 | 2T | 50% |
| Divide by 3 | 0 to 1.5 (p from 0 to 1, n from 0.5 to 1.5) | 1.5 to 3 | 3T | 50% |
| Divide by 1.5 | 0 to 1 (a), then 1.5 to 2.5 (b) | 1 to 1.5, then 2.5 to 3 | 1.5T | 67% |
In silicon, the OR gate on a clock output must be glitch-free: the two pulses here never change at the same instant, but the gate should be a characterized clock cell and the divided clock should be defined in SDC as a generated clock. Each design also depends on the input clock duty cycle, since the negative edge sets half of the output timing.
5. What are multicycle paths and false paths? How do you resolve them?
A multicycle path is a path that, by design, is allowed more than one clock cycle to propagate, for example because the capturing register is enabled only every second cycle. A false path is a path that can exist structurally but is never exercised in a way that matters, such as between asynchronous clock domains handled by a synchronizer, or between two mode-select settings that are never active together.
They are "resolved" by telling the timing tool, not by changing the logic: set_multicycle_path for multicycle paths (setup with N, and usually hold with N−1 so the hold check moves back to the right edge) and set_false_path for false paths. Each exception must be justified from the design and reviewed, because a wrong exception hides a real failure. Where possible, verify multicycle behavior with assertions or formal checks on the enable.
6. A flop to flop path with combinational delay, where the output of the second flop feeds back into the combinational logic. Which is the fastest path for a hold violation, and how will you resolve it?
Hold is checked on the shortest path. For background, see All About Clock Skew and Short Paths: Hold Violations Explained on this blog. Here there are two paths into the second flop: from flop 1 through the logic, and from flop 2's own output back through the logic to its input. The feedback path is usually the hold risk: launch and capture are the same flop, so there is no skew to help, and if the feedback enters the logic close to the D input it passes through very few gates. If flop 1 has a large positive skew to flop 2 (capture clock later), the forward path can also fail.
To fix it, add delay (buffers or delay cells) on the short path near the D pin of the capturing flop, sized so the setup slack of the longer paths sharing that pin is not consumed. Fix hold after CTS, when real skew is known, and check at the fast corner.
7. What RTL coding styles yield an optimal backend design?
- Register module outputs (and ideally inputs) so timing budgets between blocks are simple.
- Avoid long combinational chains and deep priority if-else trees; use case statements or parallel logic where priority is not needed.
- Use enable-style clock gating that synthesis can map to integrated clock-gating cells; avoid hand-built gated clocks.
- Keep a single, clean reset strategy and synchronize reset release.
- Avoid latches, combinational loops and asynchronous logic outside approved cells.
- Partition by clock and power domain so floorplanning and UPF power intent follow the hierarchy.
- Keep high-fanout control signals few and registered, so they can be buffered or replicated.
8. Draw timing diagrams to represent propagation delay, setup, hold, recovery, removal and minimum pulse width.
| Parameter | Measured from and to | What it protects |
|---|---|---|
| Propagation delay | Input change (or clock edge) to output change, usually at 50% points | Path timing |
| Setup time | Data stable before the active clock edge | Correct capture of the new value |
| Hold time | Data stable after the active clock edge | Old value not corrupted by the next one |
| Recovery time | Asynchronous reset or set deasserted before the clock edge | Clean exit from reset (the setup-like check) |
| Removal time | Asynchronous reset or set held after the clock edge | Clean exit from reset (the hold-like check) |
| Minimum pulse width | Shortest high or low time on a clock or async pin | Reliable switching of the cell |
When drawing these, show the clock edge as the reference, mark setup to the left and hold to the right, and draw recovery and removal the same way around the reset deassertion edge.
FAQ
Are these questions still relevant for SK hynix interviews?
The topics are standard for physical design and timing roles at memory and logic companies alike. Expect extra questions on current topics such as multi-patterning and advanced packaging at recent nodes.
Why does divide by 3 need the negative edge?
With only positive edges, a divide by 3 output can be high for one or two input periods, giving 33% or 67% duty cycle. The 1.5-period high time needed for 50% requires an edge in the middle of an input period.
Is a multicycle path a timing violation being hidden?
Only if it is applied wrongly. A correct multicycle constraint describes how the design really works; an incorrect one hides a real failure, which is why exceptions are reviewed.
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.