Skip to main content

AcceleratorComparison

Optimization scenarios · Implementation: mlsysim/engine/scenario_definitions/optimization.py.

Construct this scenario with keyword inputs. The fields below show the types and defaults declared by this class.

candidate against baseline: peak, power, and peak per watt.

Peak rates compare at one convention (dense FP16 unless precision says otherwise); a ratio across precisions mixes a format change into the hardware change.

Notes

Equation: Use the dimensional products, ratios, or selection rules documented on the calculated properties below.

Assumptions: The same arithmetic workload runs at comparable utilization and selected precision; rated TDP supplies the power comparison.

Excludes: Memory limits, communication, and measured workload power are absent.

References

Derivation: the calculated properties below state their defining arithmetic and units. They are scenario accounting or input-selection rules, not independently measured relationships.

Examples

>>> from mlsysim import Scenarios
>>> case = Scenarios.V100ToB200
>>> round(case.peak_ratio, 4)
18.0

Attributes

Name Description
baseline
baseline_device_hours
candidate
candidate_device_hours Candidate device-hours corresponding to the scenario’s baseline budget. Equation:
efficiency_ratio Relative peak arithmetic work per rated watt. Equation: peak_ratio / tdp_ratio.
peak_ratio Candidate peak arithmetic rate divided by baseline peak rate at the selected precision.
precision
tdp_ratio Candidate rated power relative to the baseline. Equation: ``candidate.tdp /

Methods

Name Description
equivalent_hours Candidate device-hours for the same ideal arithmetic work. Equation: ``baseline_hours /

equivalent_hours

engine.scenario_types.AcceleratorComparison.equivalent_hours(baseline_hours)

Candidate device-hours for the same ideal arithmetic work. Equation: baseline_hours / peak_ratio. Units: the input’s hour unit; assumes equal utilization at the compared peaks.

Candidate device-hours that match baseline_hours at equal sustained share of peak.

baseline

Type: HardwareNode.

Declared default: required.

candidate

Type: HardwareNode.

Declared default: required.

precision

Type: Optional[str].

Declared default: None.

baseline_device_hours

Type: Optional[float].

Declared default: None.

peak_ratio

engine.scenario_types.AcceleratorComparison.peak_ratio

Candidate peak arithmetic rate divided by baseline peak rate at the selected precision. Equation: _peak(candidate) / _peak(baseline). Units: dimensionless; achieved efficiency is not modeled.

tdp_ratio

engine.scenario_types.AcceleratorComparison.tdp_ratio

Candidate rated power relative to the baseline. Equation: candidate.tdp / baseline.tdp. Units: dimensionless; TDP is not measured workload power.

efficiency_ratio

engine.scenario_types.AcceleratorComparison.efficiency_ratio

Relative peak arithmetic work per rated watt. Equation: peak_ratio / tdp_ratio. Units: dimensionless; neither memory limits nor utilization are included.

candidate_device_hours

engine.scenario_types.AcceleratorComparison.candidate_device_hours

Candidate device-hours corresponding to the scenario’s baseline budget. Equation: equivalent_hours(baseline_device_hours). Units: device-hours under ideal peak-rate scaling.

Candidate device-hours for baseline_device_hours at equal sustained share of peak.

Back to top