Bandwidth & Latency Bottlenecks: Debug Playbook

Debug playbook

Debug Playbook for Bandwidth & Latency Bottlenecks centers on roofline position, p99 memory latency, and throughput saturation point. The objective is to connect profiler evidence to root-cause mechanism and release-safe action.

GPU debug is mechanism-first. Lock revisions, isolate first failing workload, then prove one hypothesis before applying broad tuning.

Root-cause tree

ROOT-CAUSE TREE — Bandwidth & Latency Bottlenecks

roofline position, p99 memory latency, and throughput saturation point regressed
        |
  reproducible on replay?
      /              \
    no                yes
    |                  |
env/test noise    counter triage
                   |
             compute-bound or memory-bound?
                /                  \
             compute            memory/interconnect
             issue stalls       cache/NoC/DRAM stalls

Stop at first failing mechanism, then patch.

1. Freeze workload seed, compiler, driver, firmware, and hardware tags.

2. Identify first failing metric and where it appears in timeline.

3. Classify bottleneck domain: scheduler, execution, memory, interconnect, or thermal.

4. Build one focused reproducer that isolates dominant mechanism.

5. Patch smallest owner-controlled fix.

6. Re-run quality, performance, and stability matrix.

Review memo template

GPU DESIGN REVIEW MEMO - Compute Fabric & Memory System / Bandwidth & Latency Bottlenecks

1. Symptom
   - Watched metric: roofline position, p99 memory latency, and throughput saturation point
   - Failing workload or scene: <name>
   - Impacted stage: <warp scheduling, memory hierarchy, graphics stage, interconnect>
   - Revision tags: <kernel/driver/compiler/firmware/hardware>

2. Mechanism hypothesis
   - Primary mechanism: Performance collapses when any stage (SM issue, cache, NoC, controller, link) saturates; bottleneck localization needs cross-stack correlation.
   - Competing hypotheses: <divergence, memory coalescing, scheduling, thermal throttling>
   - Missing evidence: <counter capture, trace, topology heatmap, timing report>

3. Proposed action
   - Minimal reversible change: <kernel/config/RTL/policy update>
   - Expected movement: <throughput, frame-time tail, perf-per-watt>
   - Regression risk: scheduler fairness, cache contention, thermal behavior, software compatibility

4. Signoff
   - Re-run artifact: roofline plot, bottleneck decision tree, and counter correlation dashboard
   - Required owners: performance lead, GPU architect, system integration owner
   - Final decision: ship, bounded rollout, rollback, or escalate

GPU deep dive

Fabric and memory-controller behavior decides scaling long before peak ALU utilization is reached.

Concept diagram

COMPUTE FABRIC

SM clusters <-> NoC <-> L2 <-> memory controllers <-> HBM

Metric graph

SCALING EFFICIENCY

single GPU        ███████████ 100%
with heavy NoC    ████████
with tuned QoS    █████████

Reports and artifacts

• NoC congestion map

• HBM controller queue stats

• PCIe/DMA overlap timeline

• roofline position report

Mini case study

Crossbar arbitration favored bulk traffic and starved latency-sensitive queues, collapsing tail performance.

Debug branches

• Track per-link hotspots, not only aggregate BW

• Inspect controller page-hit and queue depth behavior

• Validate host-device overlap during peak traffic windows

Senior review question

Ask: which metric and benchmark pairing proves this topic is truly closed in production context?

Key takeaways

• Always pair micro-kernel metrics with end-to-end workload impact.

• Lock toolchain, driver, and launch metadata before comparing performance results.

Common pitfalls

• Optimizing occupancy without checking memory-system saturation.

• Comparing profiler captures from different driver or compiler builds.

• Declaring wins without reproducible accuracy and performance gates.

Principal GPU review addendum

Bandwidth & Latency Bottlenecks is not just a definition to memorize. In a real GPU program it becomes an interaction between software shape, compiler mapping, warp execution, memory movement, interconnect policy, and physical limits. The first senior move is to name which layer is being exercised before interpreting a counter.

Performance collapses when any stage (SM issue, cache, NoC, controller, link) saturates; bottleneck localization needs cross-stack correlation. This mechanism matters because GPUs are throughput machines: a small inefficiency repeated across lanes, warps, SMs, frames, or dispatches can dominate product performance even when a unit-level diagram looks balanced.

Use roofline position, p99 memory latency, and throughput saturation point as an entry point, not as the conclusion. A metric shift only becomes actionable after it is tied to a workload slice, a profiler capture, an architectural path, and a reproducible artifact such as roofline plot, bottleneck decision tree, and counter correlation dashboard.

Interconnect architecture decides whether multi-SM and multi-die resources behave as one system. The review posture is therefore evidence-first: explain what the kernel or graphics workload asked for, how the GPU mapped it onto hardware, where useful work stopped, and which owner can change the smallest boundary safely.

In review, insist on a concrete chain from workload to hardware behavior: workload shape -> compiler/runtime mapping -> warp or pipeline behavior -> memory/fabric pressure -> measured product impact. That chain prevents generic GPU tuning advice from replacing engineering evidence.