Skip to content

Add action-based analytical power modeling - #104

Open
YuxuanHan0326 wants to merge 4 commits into
yx/upstream-analytical-latency-v1from
yx/upstream-power-v1
Open

Add action-based analytical power modeling#104
YuxuanHan0326 wants to merge 4 commits into
yx/upstream-analytical-latency-v1from
yx/upstream-power-v1

Conversation

@YuxuanHan0326

Copy link
Copy Markdown
Collaborator

Summary

Adds an action-based analytical power model on top of compiler CostTrace and HBM V4 traffic.

Main changes

  • Action-family dynamic-energy accounting
  • Matrix, Vector, Scalar, control, SRAM, and HBM-controller breakdowns
  • Ideal hierarchical clock-gating estimate
  • Ungated clock-power upper bound
  • External HBM3E traffic and background energy
  • Empirical low/nominal/high activity envelope
  • Versioned calibration and physical-property inputs

Validation

  • CostTrace action-family coverage: 100%
  • Main-compatible power holdout median error: 7.47%
  • Main-compatible power holdout P95 error: 22.13%
  • Detailed and compressed traces produce identical action totals
  • Ideal clock energy never exceeds the ungated result

Dependencies

This is a stacked PR. Its base is the analytical-latency branch so reviewers initially see only the Power changes.

Scope

This model excludes multi-chip links, package, cooling, CTS, SRAM leakage, and signoff-level physical effects.

Compare dense and routed-MoE detailed schedules with affine summaries across opcode counts, DMA multiplicity, compute timing, HBM service, physical traffic, and action energy.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant