Frequency spec by channel loss

[bapivee]

Change spec to specify frequency vs loss instead of distance.

[bapivee]

Or perhaps add an interposer-based trace length max to the substrate.

[kenpoulton]

Loss is probably the right metric for terminated lines. But it's not even clear to me what it means in the context of unterminated lines. I think the unterminated case should be considered as RC circuits. But I have not researched what other unterminated standards do.

We did some R+C estimates for laminate and interposer packages. For laminate, the R is so small that you can consider it to be C only. For interposer, R is important also.

For a 7+2+7 laminate, you can only conveniently do 3 signal layers in G-S-G-S-G-S-G stackup, so we considered a 3-deep-stacked BoW. We estimated 5 mm longest lines for 3-stack BoW. 5 mm lines area about 1 pF and 2 ohms. 10 mm lines gives margin for flexibility in layout and design rules and supporting a 4-stack, and has twice as much C and R: 2 pF and 4 ohms.

The R is small enough to ignore. The C drives transmitter power - I get ~0.5 pJ/bit just for driving Cwire, and I guess the pre-driver stages may take as much. Note that the power for driving Cwire is low if your actual line lengths are well under the max, but the pre-driver power is baked into the chip and doesn't depend on the wire length. So that argues that we don't want to support more than 10 mm in laminate.

In interposer with 55-um-pitch bumps, I got 1.4 mm for a 4-stack. If we allow up to 3 mm lines in interposer (for flexibility in interposer and bump layouts) then the wires are about 1.5 pF and 30 ohms. In this case the R is significant: the RC of the line by itself is 45 ps, which is close to the limit for clean RC-limited waveforms at 200 ps UI.

One way to spec this would be 2 pF max Cline and 45 ps RC product. This corresponds to about 10 mm in laminate and 3 mm in interposer, but the spec would be the C and RC numbers.