Setup: Illustrative display DMA experiment: a 64-bit AXI read channel at 800 MHz measures 4.2 GB/s against a 7 GB/s target. The 64-bit ceiling is 6.4 GB/s, so the target already requires a wider or faster path. Assume fixed full-width traffic, sufficient downstream memory capacity and otherwise unchanged conditions when comparing 64-bit and 128-bit variants.
Initial 64-bit measurement: 4.2 / 6.4 = 65.625% full-width accepted read beats. This measures utilisation, not a unique root cause.
The 128-bit variant has a 12.8 GB/s ceiling but measures 4.5 GB/s: a 7.1% throughput gain and about 35.16% accepted-beat utilisation. Wider data alone has not filled the path.
In this illustrative trace, the DMA can track 16 requests, but the controller admits only four. During the observed gaps, those four slots are occupied and new requests wait for completion. Check service latency and other limits before attributing all lost throughput to this admission cap.
A controlled replay raising controller admission to 16 measures 5.8 GB/s at 64 bits and 7.2 GB/s at 128 bits. These are illustrative observations, not values derived from outstanding depth alone; both stay below their respective data-channel ceilings.
The repaired 64-bit variant remains below the 7 GB/s target even at its 6.4 GB/s theoretical ceiling. The repaired 128-bit variant meets the bandwidth target in this replay; latency and competing workloads still need review.
Outcome: Admission capacity and data width both mattered. Repairing admission improved utilisation, while the target still required more than the original 64-bit path could carry.
Lesson: Check the physical ceiling first, then use controlled experiments to identify which additional limits prevent reaching it.
Setup: Illustrative camera replay offers 1.2 GB/s to a memory path whose long-run measured service rate is 1.5 GB/s. A 1 MB ingress buffer still overflows during a measured 2 ms interval in which this path accepts no camera data.
During the 2 ms service pause, the offered stream adds 1.2 GB/s × 0.002 s = 2.4 MB of backlog. Starting empty, a 1 MB buffer is short by 1.4 MB before service resumes.
The 1.5 GB/s long-run service measurement does not bound the longest service gap. Record accepted camera bytes and buffer occupancy over the same time axis; an average can hide a deadline or capacity failure.
Correlate the pause with arbitration grants, outstanding occupancy, return backpressure, controller events and competing traffic. A periodic correlation is evidence for investigation, not proof that refresh or any one mechanism caused the pause.
Candidate repairs are measurable alternatives: reduce the longest service gap, reserve service for the stream, increase usable buffering, or reduce the offered rate. Replay the same trace after changing one control at a time.
Outcome: The failure is established as a 1.4 MB capacity deficit under the observed service gap. Its root cause remains open until evidence distinguishes arbitration, downstream service, backpressure and other candidates.
Lesson: For continuous streams, budget the maximum backlog over a stated interval, not only long-run average bandwidth.