Skip to main content
ModulesPerformance

Power & Thermal Awareness — From Activity to perf/W

Translate simulated activity into power/thermal behavior and communicate perf/W trade-offs credibly using McPAT and HotSpot

expertPerformance140m
4
Exercises
5
Tools
4
Applications
2
Min Read

Practical Exercises

  • McPAT power model calibration against silicon data
  • HotSpot thermal simulation and floorplanning
  • DVFS policy impact analysis
  • Junction temperature vs performance trade-off study

Tools Required

McPATHotSpotPower measurement toolsThermal simulationDVFS controls

Real-World Applications

  • Processor thermal design optimization
  • Mobile SoC power budget allocation
  • Datacenter cooling requirement planning
  • Energy-efficient system design

Power & Thermal Awareness — From Activity to perf/W

Goal: Translate simulated activity into power/thermal behavior; communicate perf/W trade‑offs credibly.


📋 Table of Contents


1) Power basics

Dynamic power: P_dyn ≈ α · C · V^2 · f (α: activity factor).
Leakage: rises with temperature and process; DVFS changes both.
Memory power: DRAM self‑refresh/activate/precharge/read/write energies often dominate memory‑bound workloads.


2) Workflow (McPAT + HotSpot)

  1. Export per‑unit activity from your simulator (ops/acc, toggles, misses).
  2. Feed McPAT with structural params (tech node, pipeline, cache sizes/assoc, ports). Calibrate using a silicon anchor.
  3. Generate power traces per time‑step (e.g., every 100 µs).
  4. Build a floorplan (areas, adjacency) and feed power into HotSpot to simulate temperature (steady‑state & transient).
  5. Iterate with DVFS, clock/power gating, residency policies; observe perf/W and thermal headroom.

2.1 Calibration tips

  • Match unit‑wise power to known silicon at 1–2 representative workloads.
  • Adjust energy/operation tables within realistic bounds; document deltas.
  • Validate leakage vs. temperature curves qualitatively against public data.

3) Communicating perf/W

  • Present performance at fixed power or power at fixed performance; both are useful.
  • Show junction temperatures and time‑to‑throttle under realistic cooling.
  • Attribute power by core vs. uncore vs. memory.

4) Example artifacts

McPAT fragment (illustrative):

<processor technology="5" clock_rate="3500" voltage="0.9">
  <core id="0" pipeline_depth="14" issue_width="4" />
  <L1D size="64kB" assoc="8"/>
  <L2 size="1MB" assoc="8" banks="2"/>
</processor>

HotSpot config keys:

ambient_temp = 45
heat_sink = 0.5  # K/W
chip_thickness = 0.5e-3

Run steady‑state and transient; compare max‑T vs. throttling thresholds.


5) Caveats

  • Ignoring package/VRM and cooling assumptions leads to over‑optimistic results.
  • DRAM and I/O power can dominate; include them.
  • Model leakage‑temperature feedback; HotSpot transient runs help.

References

  • McPAT (MICRO'09) and repo; HotSpot manuals/TR.
#power#thermal#McPAT#HotSpot#DVFS#perf-per-watt#energy-efficiency