[Arm-netbook] HDMI High-Frequency Layout: Timing

[Richard Wilbur]

On Tue, Aug 1, 2017 at 4:29 AM, Luke Kenneth Casson Leighton
<lkcl at lkcl.net> wrote:
> ok reading in full, cutting extraneous, answering only with confirmation.

Great plan!

> On Tue, Aug 1, 2017 at 10:57 AM, Richard Wilbur
> <richard.wilbur at gmail.com> wrote:
>
>> [...] until we get to 4.3 inches from the
>> signal source the line driver should be able to squash most
>> reflections on the leading edges (first quarter wavelength).
>
>  actual distance is 55mm, under half the 4.3in.

I figured this wasn't going to be a problem and so I mentioned it to
ease the tenor of the discussion.

>> Inter-pair skew:  (Requirements for HDMI v1.4)
>>     clock-data skew:  Δt <150ps[3]
>>      => Δl < v * Δt = 22mm ~= 870mil
>>     T(Pixel) = 1/(pixel clock) = 2.94ns
>>     Skew(Inter-Pair) < 0.20 * T(Pixel)[4][5] = 588ps
>>      => Δl < v * Δt = 88.2mm ~= 3470mil
>>     Chrontel suggests matching between any two pairs be within 100mil.[5]
>>     => Δl < 100mil => Δt < Δl / v = 2540um / (150um/ps) = 17ps
>
>  actual difference between CK and Tx2 is 55 - 48mm, or 7mm. so... 275
> mil.  whoops.
>
>  between CK and Tx2 is 55 - 52 = 3mm, so... 118 mil.  again whoops.

I wouldn't sweat too much breaking the 100mil target by hitting 275mil.

I didn't quite organize that very well.  Here's what it should look
like.  Notice that requirements for the standard (HDMI v1.4, attested
by both Chrontel and TI) are first and suggestions are specifically
introduced with the verb, "suggests":

Inter-pair skew:  (Requirements for HDMI v1.4)
    Chrontel, TI:  T(Pixel) = 1/(pixel clock) = 2.94ns
    Skew(Inter-Pair) < 0.20 * T(Pixel)[4][5] = 588ps
     => Δl < v * Δt = 88.2mm ~= 3470mil
    Toradex suggests clock-data skew:  Δt <150ps[3]
     => Δl < v * Δt = 22mm ~= 870mil
    Chrontel suggests matching between any two pairs be within 100mil.[5]
    => Δl < 100mil => Δt < Δl / v = 2540um / (150um/ps) = 17ps

Intra-pair skew:  (Requirements for HDMI v1.4)
    Chrontel, TI:  T(bit) = 0.1 * T(Pixel) = 294ps
    Skew(Intra-Pair) = 0.15 * T(bit)[4][5] = 44.1ps
     => Δl = v * Δt =  6.62mm ~= 261 mil
    Toradex suggests:  Δt < 5ps[3]
     => Δl < v * Δt = 0.75mm ~= 30 mil
    (Chrontel suggests, without saying why, that matching between
signals should be within 5mil.[5]  Given the context and calculations
above suggesting 261 mil for total, this should probably be 5mil skew
per segment.)

>> A Texas Instruments (TI) employee
>> specifically suggested to keep the clock pair longer than the data
>> pairs.[7]
>
>  sounds like a good idea... and has to happen anyway: the clock lines
> have slightly further to go.

Happy coincidence!

>> Intra-pair skew:  (Requirements for HDMI v1.4)
[...]
>>     (Chrontel suggests, without saying why, that matching between
>> signals should be within 5mil.[5]  Given the context and calculations
>> above suggesting 261 mil for total, this should probably be 5mil skew
>> per segment.)
>
>  i try to meet that - 5 didn't know it was as little as 5mil though.
> that's absolutely tiny!

I'm going to guess that is probably due to the geometry of the
differential traces where if the spacing is 5mil then a 5mil
intra-pair skew lengthens the wavefront by sqrt(2) and it is now
turned 45 degrees from the intended direction of propagation.

>> Chamferred Corners (or Trace Bend Geometry)
>>
>> I'm glad to see you are already using 45 degree bends instead of 90
>> degree corners.  This helps the corners maintain the proper impedance.
>> When serpentine traces (meanders) are needed to attain certain lengths
>> of a single-ended trace, make sure individual segment lengths are at
>> least 1.5x the width of the trace.  Also, the spacing between parallel
>> segments of the same trace should be at least 4x the width of the
>> trace.[9]
>
>  that's going to be very very hard to achieve: there is an extremely
> limited amount of space.

This section is mostly here to talk about the 45 degree versus 90
degree bends which is important for any high-frequency trace, be it
single-ended or differential.  The last part about proportions for
meanders concerns single-ended traces (like if you needed it for CEC
or some other *non-differential* signal).  The proportions for
differential meanders are down in the next section.

>> Length Matching
>>
>> Differential pair signals should not propagate asynchronously over a
>> distance greater than 15mm[10] = 590mil.  Thus the compensation for
>> length mis-matches should be placed as close to the mismatch as
>> possible.  Differential traces can be segmented by a connector, pad
>> (component or IC), or via.[10][5]  "Each segment of a differential
>> pair connection needs to be matched individually."[10]
>
>  yeah i saw that in the toradex recommendations, otherwise there's
> skew between traces.
>
>> Ideal serpentine trace geometry for equalizing differential traces
>> consists of the following proportions:  the spacing between traces in
>> the meanders should not exceed twice the normal spacing, and the
>> length of more widely spaced traces should not exceed three times the
>> normal trace width.[10, see Figure 23]
>
>  there's a lot of other stuff in here which is really good, such as
> making sure that lengths on each *layer* are matched, and that even
> when turning corners the lengths are matched.  and matching just after
> VIAs *not* before... damn

I'm glad you're seeing good things.  Matching should happen on the
same side of the via (same segment) as the skew happens.

>  so thank you - much to correct and think about.  really appreciated
> you finding all this stuff richard.

I wanted to present the parameters and principles so that you can make
more well-informed choices.  You don't have errors to correct but
better information on which to base your choices.