AI for VLSI · All levels

Dataflow Mapping Strategies

Accelerator Architecture: Weight-stationary, output-stationary, and row-stationary mappings trade local reuse against bandwidth and scheduling complexity.

What this topic teaches

Dataflow Mapping Strategies turns AI concepts into VLSI-ready engineering decisions. Weight-stationary, output-stationary, and row-stationary mappings trade local reuse against bandwidth and scheduling complexity. The practical challenge is proving value with reproducible evidence, bounded risk, and explicit ownership.

The senior-engineer question

When weight/input/output reuse factor, buffer hit rate, and latency moves, can you identify the failing layer, the mechanism, the artifact, and the owner who can close risk with a measurable fix?

diagram

AI-VLSI FLOW — Dataflow Mapping Strategies

problem framing
      |
      v
data + model definition
      |
      v
training / optimization
      |
      v
compute-hardware mapping
      |
      v
deployment + validation

Primary metric: weight/input/output reuse factor, buffer hit rate, and latency

Picture the system

Start each review with an architecture sketch before opening dashboards. These diagrams are designed for design reviews and interview whiteboards.

Dataflow options

diagram

DATAFLOW STRATEGIES

weight-stationary : keeps weights local
output-stationary : keeps partial sums local
row-stationary    : balances multi-dimensional reuse

Tensor and data path

diagram

TENSOR / PIPELINE MAP — Dataflow Mapping Strategies

feature source -> preprocessing -> tensorized input
      |                             |
      +---- shape + scale checks ---+
                    |
                    v
            model execution / inference

Shape and scaling discipline decides correctness and portability.

Training and update loop

diagram

TRAINING TIMELINE — Dataflow Mapping Strategies

time --->
data batch      __/--/--/--/--/--/--/--
forward pass    ____/--/--/--/--/--/---
backward pass   ________/--/--/--/-----
optimizer step  ____________/--/--/----
eval checkpoint _____________/--/-------

Convergence depends on stable loop timing and signal quality.

Compute limit lens

diagram

ROOFLINE LENS — Dataflow Mapping Strategies

performance
   ^
   |                compute bound region
   |               /
   |              /
   |-------------/---------------- memory bound region
   +----------------------------------------------> operational intensity

Use this to decide compute optimization vs memory optimization.

Ownership layers

diagram

AI-VLSI OWNERSHIP LAYERS — Dataflow Mapping Strategies

layer                    owns                               typical failure
---------------------    --------------------------------   ----------------------------
problem framing          metric + acceptance criteria       wrong objective target
model + training         representation + optimization      unstable or biased model
hardware mapping         dataflow + memory + precision      bandwidth stalls / mismatch
deployment stack         runtime + firmware + drivers       latency jitter / incompatibility
governance               monitoring + rollback + signoff    silent drift in production

Evidence to collect

Primary metric: weight/input/output reuse factor, buffer hit rate, and latency.
Primary artifact: dataflow comparison matrix, mapping trace, and reuse analysis.
Owners to bring into review: compiler engineer, architecture owner, performance analyst.
One workload slice where behavior regressed and one where it held.
One profile view that separates model issue from runtime/hardware issue.

Ownership map

diagram

OWNERSHIP MAP — Dataflow Mapping Strategies

artifact focus         owner
------------------     ----------------------------
modeling          compiler engineer
architecture      architecture owner
integration       performance analyst

Production issues happen when ownership is assumed, not declared.

Subpages in this topic

Each topic is taught across mechanism, inputs/outputs, reports, debug, worked example, pitfalls, interview, checklist, theory, design space, expanded case study, walkthrough, comparison matrix, software view, and silicon impact.

Key takeaways

Always map ML metrics to engineering decisions and release risk.
Separate data/model issues from hardware/runtime bottlenecks before fixing.
Use reproducible artifacts and owner signoff for every rollout decision.

Common pitfalls

Benchmark wins with no signoff correlation.
Ignoring calibration and drift when deploying quantized models.
Shipping without a rollback and ownership matrix.

AI-VLSI deep dive

Dataflow and memory hierarchy choices dominate practical accelerator performance.

Concept diagram

diagram

ACCELERATOR STACK

MAC array -> dataflow map -> SRAM hierarchy -> NoC/DRAM

Metric graph

diagram

BOTTLENECK MIX

array underutilization ████
SRAM pressure          ███████
NoC/DRAM congestion    ██████

Reports and artifacts

array utilization
reuse factor report
buffer hit analysis
NoC/DRAM bottleneck trace

Mini case study

A dataflow change improved reuse but exposed NoC congestion, requiring route-policy updates.

Debug branches

Map reuse factors
Inspect memory stall sources
Profile NoC arbitration

Senior review question

Ask: what evidence connects this ML claim to a concrete VLSI workflow decision and owner signoff?

Key takeaways

Every AI claim should map to a measurable engineering outcome.
Validate both model quality and hardware/runtime feasibility before adoption.

Common pitfalls

Optimizing benchmark metrics that do not correlate with signoff goals.
Ignoring data drift and calibration after deployment.
Shipping ML workflows without clear rollback ownership.

Execution drill pack 1

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 1

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 2

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 2

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 3

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 3

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 4

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 4

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 5

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 5

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 6

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 6

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Execution drill pack 7

Use this pack to rehearse AI-for-VLSI decision making on ai-vlsi/accelerator-architecture/dataflow-mapping-strategies: metric framing, mechanism proof, hardware implications, and release safety.

Evidence checklist

Metric context includes workload, dataset slice, and revision tags.
Mechanism explanation links model behavior to observed outcome.
Hardware/runtime feasibility is profiled, not assumed.
Owner and rollback path are documented before rollout.

Review prompts

Which decision will this model output influence?
What is the first failing layer when metric regresses?
Which owner applies the smallest reversible fix?
What validation matrix is required before deployment?

Evidence capsule

diagram

AI-VLSI EVIDENCE CAPSULE 7

PATH: ai-vlsi/accelerator-architecture/dataflow-mapping-strategies
WORKLOAD SLICE: <name>
PRIMARY METRIC: <value/trend>
FIRST FAILING LAYER: <data/model/runtime/hardware>
OWNER: <name>
PRIMARY ARTIFACT: <report/profile/dashboard>
DECISION: <ship / rollback / escalate>

Practice this lesson