AVR Freaks

It's not just Freaks - pretty much universal to any microcontroller!

Personally I've never seen a reason to omit a 20 cent crystal

indeed! 

Over what temperature range will your equipment need to operate? 

If always in-doors in normal room conditions, it "may" work, over extend temperature range, may not with out a xtal!

Jim  I have no experience with the new xtiny's!

As your DMX signal is always present why not use it to run-time calibrate your controller?

first data byte always 0,

Until it isn't.  If you try to run your DMX receiver on an RDM universe (or a DMX universe which uses vendor-specific start codes for things like dimmer pack configuration), you may get a nasty surprise.

Spaces will always (except for the break) come in bit widths of 1-to-9 bits.  Better to periodically time a range of spaces in a frame (even all of them, but excluding the break), and bin them into 1-bit, 2-bit, etc bins.  At +/-3%, you'll never confuse bins.  Then, compute the average length of each bin, and then the average length of one bit.

At 16 MHz, a bit width is 64 cycles.  Fairly easy to use input capture to time spaces, but you may need to resort to assembly instead of C for the timing-critical parts.  If the average is less than 64 bits, your clock is slower than the lighting desk.  If the average is more than 64 bits, your clock is faster than the lighting desk.  Tweak the oscillator calibration accordingly to target an exactly 64-cycle average bit length.

EDIT: You can have as many as 5 spaces per slot (start bit, and data bits 1, 3, 5, and 7 as '0', for a value of 0x55), or as few as 1 space per slot (start bit followed by 8 '0' data bits).  With a 512-address universe, plus the start code, you can have as many as 2,565 spaces, or as few as 513.  With each 1-bit space having a nominal width of 64 cycles, the cumulative count in bin #1 could reach as high as 164,160, requiring a 3-byte accumulator.  At the other end, each 9-bit space will have a nominal width of 576 cycles, so the cumulative count in bin #9 could reach as high as 295,488, so each bin will require a 3-byte accumulator.  You'd also need to maintain a 2-byte counter for each bin.  Total memory cost would be 45 bytes for the bin states.  Mind you, this is if you decide to profile an entire frame.  You could limit your captures to just a handful of slots and that would likely be enough.

...in response to ...

Now, in #19 jgm mentioned the design/app criteria a bit.  Indeed, that is what I'll say in response:  As it [almost?] always does, "It depends".

Is OP making a handful of these devices, for personal use?  Then even with "free" time OP may be way beyond the payback.  [I realize this is pre-build discussion]

Now, I have some production apps over the years where we have chosen internal oscillator, though it has a UART link.  With datasheet examination and experimentation, the internal oscillator is found to stay pretty flat at constant supplyV and temperature.  [YMMV]  The comms link, in all likelihood, will be used in a benign [temperature] environment.  So we do a one-time cal at ISP time.  No hit to our product reputation over the years of production. [n.b. the driver behind this is the much faster wakeup time from powerdonw; a crystal takes some milliseconds and also draws some power]

It depends.  One size does not fit all.  In general, I agree with the suggestion to use an accurate clock source when warranted.

mostly because of save space and reduce component quantity 1 crystall + 2 caps, less component less pick and place

Use a ceramic resonator, they are a three-terminal device with integrated caps.  Single part to pick and place.  That's what the UNO R3 uses.  Specifically, the part number is CSTCE16M0V53-R0.

Digikey lists it at $0.26858 for qty >= 1000.  Board space used is minimal, overall SMD package/footprint is 3.2mm L x 1.3mm W x 0.9mm H.  Both SMD and PTH available.  Tolerance of that particular part is +/-0.5%, sufficient for comms including DMX.

There are over 1600 parts in the DigiKey catalog, many less expensive, but about the cheapest (from them) suitable SMD part will be about $0.22/qty1000.  PTH is cheaper still.

I'm wondering how much time you have spent researching this AVR clock issue before the posted topics here?  Did you determine why you were previously "unsuccessful"? How far off was the particular failing sample from nominal?  IIRC that model line's clock should be fairly stable at steady temperature.  Did you do a one-time or continuous cal using some reference source?  [Note that while "autobauding" is generally easier with a fixed trigger byte value, surely one could find a bit or a few in the data stream to help tune.]

In any case, re the 2%, have you looked in the [wait for it] datasheet?  In this case, the Asynchronous Operational Range section of the USART chapter.  There is discussion and formulas and tables.

Nice post Joey!

Jim

I am aware of disadvantage of dasy-chain
Requirements allow it and accept consequences

I think you've misunderstood me, or you've misunderstood the DMX standard.  Re-broadcast of the incoming stream does not comply with the standard.  The 'OUT' connector on your device must be electrically in parallel with the 'IN' connector.  Your device need not, and must not, re-broadcast the data stream.  Doing so makes it a repeater rather than a passive node.  I strongly recommend against this.  Any fault in your device will affect all downstream devices.

drift caused by %2 tolerance doesnt interfere 8.9.10 sampling, I will recalculate for worst combinatin Tx +2%, Rx - 2% to see where exactly is...

Don't forget that figure is meant to accommodate real-world conditions.  I've shown you the calculations.  What you can't calculate is the degree of edge smear, noise, and signal reflections you'll have to contend with on a real installation.  Don't mess around.  Use a resonator or a crystal, or calibrate in software to make sure you are as accurate as you can be.  When calibrating, you >>must<< take an average of several spaces and/or marks, otherwise you're likely to arrive at an inaccurate measure of the real speed due to the aforementioned effects of edge smear, noise, and signal reflections.