A worked mechanism
Choose which state survives before drawing the wake sequence.
Compare state classes in the existing compute-island policy; retention is selective rather than a promise that every register survives.
Retained configuration
Save → hold → restore
mode_reg and calibration_epoch use a retention supply and the specified save/restore controls.
Ordinary pipeline state
No implicit survival
Non-retained work is drained or cancelled and reinitialized after power returns.
Always-on authority
Separate power domain
The AON manager preserves wake cause and sequencing ownership while compute is off.
Cold-off alternative
Policy change
A design may discard retained state only under a different declared policy that initializes instead of restoring it.
Work it through
Reset must not erase a restored mode
If mode_reg is intended to survive, its reset and restore precedence must agree with the retention cell and power sequence. Blindly applying a later clearing reset defeats the policy.
Verify the declared post-wake value, not just restore_done.
Watch for this
Retention requires valid retention power and cell-specific sequencing; a saved bit is not protection against loss of that supply.
Check your reasoning
Compare RTL and UPF state ownership, save/restore acknowledgement, reset precedence, and post-wake architectural values.
Check your understanding
Questions to reason through.
What is the safe order for powering down a domain?
Stop new work, drain or cancel accepted work according to the protocol, save retained state, assert isolation while source outputs are still valid, gate clocks when the domain is idle, then remove power. Wake-up reverses the dependency carefully: restore power, wait for stability, enable clocks, restore or reset state, validate readiness, and only then release isolation and accept traffic.
Put it into practice
Explain the decision behind the design.
Practice Clock, Reset, CDC and Power in the question bankComplete the guide

