AI Accelerator Design · All levels

Dataflow Architectures

How stationary styles, buffer placement, and scheduling policies shape bandwidth demand, PE utilization, and energy efficiency in accelerator execution.

Section goal

How stationary styles, buffer placement, and scheduling policies shape bandwidth demand, PE utilization, and energy efficiency in accelerator execution.

How to study this section

Start with each topic hub and restate the mechanism in your own words.
Use reports and debug pages to separate symptoms from root causes.
Practice worked examples and interview drills under fixed metadata.
Close with checklist and silicon impact before making release claims.

Topics

input-stationary-dataflow/ - Input-Stationary Dataflow
output-stationary-deep-dive/ - Output-Stationary Deep Dive
weight-stationary-deep-dive/ - Weight-Stationary Deep Dive
hybrid-dataflow-selection/ - Hybrid Dataflow Selection

AI accelerator deep dive

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

Concept diagram

diagram

DATAFLOW DECISION

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

Metric graph

diagram

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?

Practice this lesson

Navigate