AI Accelerator Design · All levels

Output-Stationary Dataflow

Systolic Arrays: Output-stationary dataflow keeps partial sums local in each PE accumulator until an output tile is fully reduced, minimizing partial-sum writeback/readback traffic. Activations and weights move through the array while each PE repeatedly accumulates contributions for its assigned output coordinate, then emits the final result once complete. This can materially reduce memory energy when reduction depth is high, but requires accumulator sizing, precision management, and overflow controls that match workload statistics. If reductions exceed local capacity or pipeline timing is misaligned, frequent spills to buffer hierarchy erase the intended efficiency gains.

What this topic teaches

Output-Stationary Dataflow converts accelerator architecture concepts into release-ready engineering decisions. Output-stationary dataflow keeps partial sums local in each PE accumulator until an output tile is fully reduced, minimizing partial-sum writeback/readback traffic. Activations and weights move through the array while each PE repeatedly accumulates contributions for its assigned output coordinate, then emits the final result once complete. This can materially reduce memory energy when reduction depth is high, but requires accumulator sizing, precision management, and overflow controls that match workload statistics. If reductions exceed local capacity or pipeline timing is misaligned, frequent spills to buffer hierarchy erase the intended efficiency gains.

Senior-engineer framing question

When Partial-sum spill rate and accumulator energy per output element across matrix and convolution workloads. regresses, can you isolate the first failing execution mechanism, collect decisive evidence, assign owners, and close with rollback-safe validation?

diagram

ACCELERATOR EXECUTION FLOW - Output-Stationary Dataflow

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 rate and accumulator energy per output element across matrix and convolution workloads..
Primary artifact: Accumulator sizing and spill analysis with precision budget and writeback frequency breakdown..
Owners to include: microarchitecture owner, numeric precision owner, compiler mapping 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

diagram

BANDWIDTH LENS - Output-Stationary Dataflow

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

Primary metric tracked:
Partial-sum spill rate and accumulator energy per output element across matrix and convolution workloads.

Ownership layers

diagram

OWNERSHIP LAYERS - Output-Stationary Dataflow

+----------------------+--------------------------------+--------------------------------+
| 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|
| compiler mapping 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

Systolic efficiency is governed by feed quality, tile fit, and bubble control.

Concept diagram

diagram

SYSTOLIC WAVEFLOW

operand stream -> wavefront launch -> PE mesh compute -> reduction/writeback
                         ^ bubbles and feed stalls reduce realized throughput

Metric graph

diagram

UTILIZATION LOSSES

tile mismatch      ███████
feed stalls        █████████
sync bubbles       █████

Metrics and artifacts to collect

mesh occupancy timeline
fill-drain overhead
tile mismatch histogram
DRAM stall attribution

Mini case study

Increasing mesh size did not help until scheduling and tile alignment removed persistent wavefront bubbles.

Debug branches

Measure bubble source first
Classify compute vs memory starvation
Tune tile policy before frequency changes

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.

Practice this lesson