Ernie sits slumped at his bench, massaging his temples, staring at the scope. The smell of stale coffee permeates the lab. Ashes from a half-smoked cigar mix with the damp, cold midnight air as they swirl across the clutter of the latest unfinished project before him.
Ernie is stumped by skew. His clock driver distributes clock to multiple receivers, all identical except line CLK8 (figure 1). Clumped at the end of line CLK8 is a conglomeration of four loads with an aggregate input capacitance of 12 pF. This large capacitive load modifies the effective delay of the trace, causing a noticeable amount of skew.
On a source-terminated transmission line, the additional delay that a capacitive load contributes-above and beyond the bulk one-way transport delay of an unloaded pc-board trace-equals approximately Z0CIN. That approximation yields the group delay of the RC filter formed by the source impedance of the series-terminated driver, Z0, driving a capacitive load, CIN.
On clock lines 1 through 7, Z0=50 Ω, and CIN=3 pF, so the additional delay on each of these lines equals 150 psec. On line CLK8, the load capacitance is four times larger, causing an additional delay of 600 psec. This difference in delay adds 450 psec to Ernie's skew budget and subtracts much more from his late-night social life.
Instead of the series termination, on CLK8 Ernie should try an end termination (Figure 1, ALT #C). Because the line is long compared with the signal rise or fall time, an end-terminated line will drive the load in a special way. On each edge, the load will initially experience a combination of the line impedance, Z0, in parallel with the end-termination impedance, also Z0, forming a combined driving-point impedance of Z0/2. On any long, end-terminated structure, the additional delay that a capacitive load contributes therefore equals only (1/2)Z0CIN, half the delay of an equivalent-series-terminated line. This reduction in delay comes at the expense of the small reflection illustrated in Figure 2. This reflection is visible in the received signal one round-trip time after the arrival of each signal edge. The reflection has an amplitude of approximately (1/2)Z0CIN/T10-90, where T10-90 is the rise or fall time of the driver.
Unfortunately, Ernie chooses instead to reduce the value of series-terminating component, R8 (FIgure 1, ALT #D). Ernie reasons that this change should induce some intentional overshoot, partially compensating for the lack of vivacity in the receiver signal and slightly speeding up the threshold crossing.
Although such reasoning is technically correct, it overlooks natural fluctuations in the value of CIN among different units in production (Figure 1, ALT#E). Because a low value of R8 creates a highly resonant circuit, Ernie may eventually discover that some particular values of CIN dangerously exacerbate the overshoot on this 2.5-V net, leading to latch-up in the receiver and strangely unpredictable values of delay. The risks associated with intentional overshoot outweigh the benefits, especially when a simple, nonresonant end termination provides an equivalent improvement.