RooflinePointScenario
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.
A workload of intensity on hardware’s roofline: attainable rate and share of peak.
A measured kernel may be placed from its counter readings instead (achieved_rate and achieved_bandwidth, added 2026-10-20): its intensity is their ratio and bandwidth_utilization its share of peak memory bandwidth.
Notes
Equation: Use the dimensional products, ratios, or selection rules documented on the calculated properties below.
Assumptions: Attainable arithmetic throughput is bounded by catalog compute peak and line-rate bandwidth times intensity.
Excludes: Dispatch overhead, reduced utilization, and cache behavior are absent from this peak bound.
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.LayerNormMeasuredH100
>>> str(case.point_intensity)
'1.7857142857142858 flop/B'Attributes
| Name | Description |
|---|---|
| achieved_bandwidth | |
| achieved_rate | |
| attainable | Roofline throughput upper bound at the point intensity. Equation: min(peak, | | [bandwidth_utilization](#mlsysim.engine.scenario_types.RooflinePointScenario.bandwidth_utilization) | Achieved memory rate as a fraction of catalog peak. Equation:achieved_bandwidth / |
| hardware | |
| intensity | |
| peak | Peak compute rate at the selected arithmetic precision. Equation: catalog default or |
| peak_fraction | Roofline upper bound relative to peak compute rate. Equation: attainable / peak. |
| point_intensity | Stated arithmetic intensity or the ratio inferred from achieved rates. Equation: |
| precision | |
| ridge | Peak Roofline crossover intensity. Equation: peak / hardware.memory.bandwidth. |
hardware
Type: HardwareNode.
Declared default: required.
intensity
Type: Optional[Quantity].
Declared default: None.
precision
Type: Optional[str].
Declared default: None.
achieved_rate
Type: Optional[Quantity].
Declared default: None.
achieved_bandwidth
Type: Optional[Quantity].
Declared default: None.
point_intensity
engine.scenario_types.RooflinePointScenario.point_intensity
Stated arithmetic intensity or the ratio inferred from achieved rates. Equation: intensity else achieved_rate / achieved_bandwidth. Units: FLOPs/byte; inferred rates must describe the same workload interval.
The stated intensity, or achieved rate over achieved bandwidth for a measured kernel.
bandwidth_utilization
engine.scenario_types.RooflinePointScenario.bandwidth_utilization
Achieved memory rate as a fraction of catalog peak. Equation: achieved_bandwidth / hardware.memory.bandwidth. Units: dimensionless; the achieved rate is an input.
peak
engine.scenario_types.RooflinePointScenario.peak
Peak compute rate at the selected arithmetic precision. Equation: catalog default or precision_flops[precision]. Units: FLOP/s; achieved efficiency is not applied.
ridge
engine.scenario_types.RooflinePointScenario.ridge
Peak Roofline crossover intensity. Equation: peak / hardware.memory.bandwidth. Units: FLOPs/byte; bandwidth is the catalog maximum.
attainable
engine.scenario_types.RooflinePointScenario.attainable
Roofline throughput upper bound at the point intensity. Equation: min(peak, hardware.memory.bandwidth * point_intensity). Units: FLOP/s; overheads and reduced utilization can lower achieved throughput.
peak_fraction
engine.scenario_types.RooflinePointScenario.peak_fraction
Roofline upper bound relative to peak compute rate. Equation: attainable / peak. Units: dimensionless; this is not measured MFU.