A worked mechanism
A FIFO needs spare service to recover after a stall.
Extend the chapter’s 12 GB/s arrival, 80-cycle outage, and 500 MHz fixture. The producer cannot be throttled during the outage.
Service stops
160 ns
80 cycles at 500 MHz is 160 ns with zero downstream service.
Backlog accumulates
60 × 32 B beats
12 GB/s × 160 ns = 1,920 bytes, before margin.
Service resumes
2 GB/s surplus
At 14 GB/s service with 12 GB/s continuing arrivals, only 2 GB/s drains backlog.
Backlog clears
480 cycles
1,920 B / 2 GB/s = 960 ns, or 480 cycles, after recovery begins.
Work it through
Why equal average rates leave no recovery margin
At exactly 12 GB/s resumed service, new arrivals consume the entire service rate. The 1,920-byte backlog never shrinks in this constant-rate model.
Bound outage frequency and recovery time as well as FIFO depth.
Watch for this
This fluid-rate estimate excludes packetization and service jitter. It is a lower-level sizing argument, not a worst-case proof by itself.
Check your reasoning
Check cumulative arrivals minus cumulative departures, downstream service guarantees, high-water mark, and the interval until the next allowed outage.
Check your understanding
Questions to reason through.
How do you know whether to add buffering or improve service rate?
Buffering absorbs finite bursts and latency variation, but it cannot repair a sustained arrival rate that exceeds sustained service rate. First compare long-term rates, then size finite storage from the maximum permitted burst and response-latency envelope with explicit backpressure assumptions.
What belongs in a microarchitecture handoff?
Include requirements traceability, block and clock/reset/power boundaries, interface contracts, state ownership, ordering and error rules, cycle-level scenarios, performance and storage calculations, configuration and status behavior, observability hooks, unresolved risks, and a verification strategy for every important invariant.
Put it into practice
Explain the decision behind the design.
- Size a FIFO under unstable rateSeparate finite burst absorption from sustained overload.→
- Budget an SoC from a workloadTranslate traffic, latency, and power intent into measurable architecture budgets.→
- Trade sharing against throughputDecide when a shared engine meets the service contract and when parallel hardware is required.→
Complete the guide

