Output-Stationary Deep Dive

What this topic teaches

Output-Stationary Deep Dive converts accelerator architecture concepts into release-ready engineering decisions. Output-stationary mapping holds partial sums close to compute until reduction completion, minimizing repeated writeback and reload of intermediate values. Activations and weights traverse the array while each PE accumulates into local state, then emits final outputs when tiles close. This can materially improve efficiency for reduction-heavy operators, but only if accumulator sizing, saturation policy, and update timing align with workload statistics. If local accumulation capacity is undersized, spill traffic to shared buffers grows quickly and can dominate both latency and power.

Senior-engineer framing question

When Partial-sum spill frequency and accumulator energy per output element at target precision modes. regresses, can you isolate the first failing execution mechanism, collect decisive evidence, assign owners, and close with rollback-safe validation?

ACCELERATOR EXECUTION FLOW - Output-Stationary Deep Dive

request ingress and model metadata
      |
      v
graph lowering and kernel selection
      |
      v
tile/dataflow scheduling and memory placement
      |
      v
tensor execution + synchronization barriers
      |
      v
result assembly + quality/SLA validation
      |
      v
release decision and rollback guardrails

Evidence to collect

• Primary metric: Partial-sum spill frequency and accumulator energy per output element at target precision modes..

• Primary artifact: Accumulator sizing and spill-risk analysis with precision policy and writeback thresholds..

• Owners to include: microarchitecture owner, numeric precision owner, verification owner, post-silicon performance owner.

• One reproducible failing workload and one stable comparator run.

• One fixed-metadata run with compiler/runtime/hardware tags locked.

Bandwidth lens

BANDWIDTH LENS - Output-Stationary Deep Dive

working-set pressure
  ^
  |                saturation zone
  |          ----------------------------
  |      o   unstable tail latency
  |   o      tuning candidate
  | o        baseline behavior
  +-------------------------------------> optimization iteration

Primary metric tracked:
Partial-sum spill frequency and accumulator energy per output element at target precision modes.

Ownership layers

OWNERSHIP LAYERS - Output-Stationary Deep Dive

+----------------------+--------------------------------+--------------------------------+
| Team                 | Primary responsibility         | Closure artifact               |
+----------------------+--------------------------------+--------------------------------+
| microarchitecture owner | mechanism and architecture intent| design rationale + tradeoffs   |
| numeric precision owner | mapping, runtime, and execution   | profile traces + bottleneck map|
| verification owner | correctness, risk, and signoff    | test report + closure memo     |
+----------------------+--------------------------------+--------------------------------+

Key takeaways

• Start with mechanism classification before changing tuning knobs.

• Use one proving artifact for each major claim in review discussions.

• Close with explicit owners, validation matrix, and rollback criteria.

Common pitfalls

• Optimizing only peak throughput while p99 latency or quality regresses.

• Mixing evidence captured from mismatched runtime or thermal conditions.

• Declaring closure without production-like replay and guardrail checks.

AI accelerator deep dive

Dataflow choices are durable architecture decisions that shape memory and scheduling cost.

Concept diagram

DATAFLOW DECISION

input/output/weight stationary
  -> locality pattern
  -> movement cost
  -> throughput and power

Metric graph

DATAFLOW COST MIX

activation traffic   ███████
weight traffic       █████
partial-sum traffic  ██████

Metrics and artifacts to collect

• reuse factor map

• buffer pressure profile

• NoC traffic mix

• shape sensitivity analysis

Mini case study

A dataflow that won for convolution lost on attention-heavy batches due to activation movement pressure.

Debug branches

• Segment by model family

• Compare reuse vs movement

• Re-check mapping assumptions under batch variance

Senior review question

Ask: which first-principles bottleneck class explains the symptom, and what artifact proves it reproducibly?

Key takeaways

• Tie every accelerator claim to a reproducible workload slice and one primary metric trend.

• Prefer bounded fixes with clear owner and rollback boundary over broad tuning bundles.

Common pitfalls

• Optimizing synthetic kernels without production-shape validation.

• Reading average latency while ignoring p95 and p99 behavior.

• Declaring sparse or precision wins without fallback and quality evidence.