6.3Clock skew, metastability and pipelining
The idea
Clock skew is a difference in clock arrival times. Metastability is a flip-flop stuck between 0 and 1 after a violated setup or hold time. Pipelining cuts a long path into stages to allow a faster clock.
Think of it like this: An assembly line: each station is quick, so a new product leaves often, though each item takes longer in total.
Why it exists
Concept questions test what each effect does and what cures it.
How it works
- Skew that delays the capturing clock helps setup but can break hold.
- Metastability happens when an asynchronous input changes inside the setup/hold window; a two-flip-flop synchroniser reduces the risk.
- Pipelining: clock period ≈ stage delay + register overhead.
- Throughput improves by about k times, latency gets slightly worse.
Worked example: 24 ns of logic in 4 stages, 2 ns overhead
Period falls from 26 ns to 8 ns: about 3.25× throughput, but latency rises from 26 ns to 32 ns.
Now play with it
- →Change the number of stages and watch the latency column.
- →Read the problem table: which fixes work for hold?
Pipelining improves throughput but the latency of a single result gets slightly worse because of the register overhead.
| Problem | Cause | Cure |
|---|---|---|
| Setup violation | data arrives too late | slower clock, faster logic, pipelining |
| Hold violation | data changes too soon after the edge | add delay to the short path, reduce skew (clock period does not help) |
| Metastability | asynchronous input changes in the setup/hold window | two-flip-flop synchroniser, handshake |
| Clock skew | clock reaches flip-flops at different times | balanced clock tree |
Check yourself
Pipelining a long combinational path mainly improves
Exam notes (the dense version)
Clock skew is the arrival-time difference of the clock at two flip-flops; skew that delays the capturing flip-flop gives extra setup time but may violate hold. Metastability occurs when an asynchronous input changes inside the setup/hold window, leaving the output at an indeterminate level for an unpredictable time; it is handled with two-flip-flop synchronisers, not removed. Pipelining inserts registers to split a long path: the clock period drops to the slowest stage plus register overhead, throughput rises, but latency grows by the register delays.