These pages of design guidelines and criteria come from NASA. They make a very good read, with a good amount of discussion on critical aspects of digital design. Start with the section on clocks:
Design Guidelines and Criteria for Space Flight Digital Electronics: Section III, Clocks
What the guidelines are
The full document is Design Guidelines and Criteria for Space Flight Digital Electronics, prepared by Richard B. Katz of the NASA Office of Logic Design and hosted on klabs.org. It is based on NASA SP-8070, Space Vehicle Design Criteria, Spaceborne Digital Computer Systems. Its purpose is to give principles and criteria for specifying, designing and reviewing digital electronics for spacecraft, so the hardware is fit to fly.
It is meant for formal design reviews (preliminary design review, critical design review and pre-ship review) and detailed peer reviews. The authors say plainly that it is not an exhaustive pass or fail checklist; it is a set of questions a reviewer should be able to answer.
Why it is worth reading even if you never build for space
Spacecraft electronics cannot be repaired after launch, and they must survive radiation, temperature swings and many years of aging. That forces a discipline every chip and FPGA designer can use: prove timing on paper for the worst case, do not trust a tool to catch everything, and document the reasoning so a reviewer can check it. The clock section in particular reads like a checklist of the bugs that cause intermittent field failures in any product.
What the clocks section covers
- Skew: do not assume general-purpose routing has zero skew; use low-skew clock resources when neighboring flip-flops share a clock edge.
- Chip-to-chip timing: check both setup and hold, and for hold, pair a slow clock path with a fast data path.
- The clock tree: check PLL and DLL lock behavior, draw one clock tree per oscillator, find every signal that crosses a clock domain, and calculate worst-case duty cycle.
- Asynchronous inputs: use proper synchronizers and calculate the metastability failure rate.
The rest of the document
Part 1 of the guidelines, the low-level design rules, has twelve sections:
- Special pins
- Input/output
- Clocks
- Finite state machines
- Reset
- Hazard analysis
- Power system
- EEPROM, flash, FRAM and other non-volatile memories
- Timing analysis and margins
- Miscellaneous design guidelines and criteria
- Design and analysis documentation
- Review of digital electronic circuits
Deep dive: the numbers and details
This part holds the figures from the clocks and timing sections and a worked hold-time check. Skip it if the overview is all you need.
Numbers from the clocks section
| Item | Guideline |
|---|---|
| Oscillator duty cycle | 50 ±10% is common; use the worst case, not 50%, when data passes between opposite clock edges |
| Hold time at the receiving chip | Aim for 0 ns or less (a negative hold requirement) for chip-to-chip paths |
| Oscillator start-up | Example test data with a 50 ms power supply rise time; check start-up and PLL lock time at power-up |
| Synchronizer | Two flip-flops in series are common, but may not be enough at high clock rates; calculate the failure rate |
| PLL and DLL | Check frequency range, lock time and behavior after a single-event upset |
Numbers from the timing analysis section
- Allow ±10% change in propagation delay over mission life (a typical program specification).
- Allow ±10% change in propagation delay from radiation.
- Memory access times can vary widely; the document cites up to 4:1 variation for some ROM and PROM parts, so screen for both best and worst case.
- Dynamic simulation at worst-case values does not guarantee worst-case timing; use min-max (extreme value) analysis.
- Delays may not track with aging, so hold margin cannot be demonstrated by test alone.
Worked example: a chip-to-chip hold check
Illustrative numbers. Chip A sends data to chip B on the same clock edge. For hold, the guideline says to pair the slowest clock path with the fastest data path.
- Clock to chip B arrives late: board clock skew up to 1.5 ns.
- Fastest data path: chip A clock-to-output minimum 1.0 ns plus board trace minimum 0.3 ns = 1.3 ns.
- Chip B hold requirement: 0.5 ns.
Hold slack = data arrival - (clock skew + hold) = 1.3 - (1.5 + 0.5) = -0.7 ns. The design fails hold. If chip B instead had a hold requirement of 0 ns or less, as the guideline recommends, the slack would be 1.3 - 1.5 = -0.2 ns: still a failure, so the skew itself must be cut, for example with a low-skew clock buffer. Now add the ±10% aging and radiation allowances to every delay and redo the check at both extremes; that is the extreme value analysis the document asks for.
Worked example: opposite-edge clocking and duty cycle
Illustrative numbers. A 50 MHz oscillator has a 20 ns period. Data launched on the rising edge and captured on the falling edge has half a period, nominally 10 ns. With a duty cycle of 50 ±10%, the high time can be as short as 40% of 20 ns = 8 ns. If the path needs 7 ns of logic plus 1.5 ns of setup, it passes at 10 ns but fails at 8 ns. This is why the clocks section asks for the worst-case duty cycle instead of an assumed 50%.
Metastability failure rate
The usual model for a synchronizer's mean time between failures is MTBF = exp(t / τ) / (T0 × f_clk × f_data), where t is the time allowed for the first flip-flop to resolve, τ and T0 are device constants, and the two frequencies are the clock and data change rates. Because resolution time appears in the exponent, each extra flip-flop stage multiplies MTBF enormously, while doubling the clock rate both shortens the resolution time and doubles the denominator. That is why the guidelines warn that two flip-flops may not be adequate at high speed: the calculation, not the habit, decides.


Thanks.
ReplyDelete