ParameterBandwidthGrowthScenario
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.
Dense-model parameter growth against per-device memory-bandwidth growth.
Each series is indexed to its baseline entry; the growth is the largest index. sparse_models are reported but kept out of the growth figure, because their total parameter count does not set per-token traffic.
Notes
Equation: The calculation members below define bandwidth_index, dense_index, sparse_index, bandwidth_growth, parameter_growth.
Assumptions: The supplied baseline is comparable to every member; dense total parameters proxy a weight read.
Excludes: Sparse active-parameter traffic and changes in numerical precision.
References
Derivation: Dense-model parameter growth against per-device memory-bandwidth growth. Each member records the corresponding algebra or named physics function.
Examples
Evaluate the named configuration and display a modeled result.
>>> from mlsysim import Scenarios
>>> case = Scenarios.DenseParametersVsBandwidthSince2014
>>> value = case.parameter_growth
>>> print(f"{value:.3g}")
3.9e+03Attributes
| Name | Description |
|---|---|
| accelerators | |
| bandwidth_growth | Largest dimensionless bandwidth index in the selected accelerator series. |
| bandwidth_index | Each accelerator’s bandwidth divided by the baseline bandwidth, dimensionless. |
| baseline_accelerator | |
| baseline_model | |
| dense_index | Each dense model’s parameter count divided by the baseline count, dimensionless. |
| dense_models | |
| parameter_growth | Largest dense-model parameter index; sparse models are excluded. |
| sparse_index | Total sparse-model parameters relative to baseline; not active parameters per token. |
| sparse_models |
accelerators
Type: List[HardwareNode].
Declared default: required.
dense_models
Type: List[Workload].
Declared default: required.
sparse_models
Type: List[Workload].
Declared default: [].
baseline_accelerator
Type: HardwareNode.
Declared default: required.
baseline_model
Type: Workload.
Declared default: required.
bandwidth_index
engine.scenario_types.ParameterBandwidthGrowthScenario.bandwidth_index
Each accelerator’s bandwidth divided by the baseline bandwidth, dimensionless.
Equation: [(h.memory.bandwidth / b0).to('').magnitude for h in accelerators].
dense_index
engine.scenario_types.ParameterBandwidthGrowthScenario.dense_index
Each dense model’s parameter count divided by the baseline count, dimensionless.
Equation: [(m.parameters / p0).to('').magnitude for m in dense_models].
sparse_index
engine.scenario_types.ParameterBandwidthGrowthScenario.sparse_index
Total sparse-model parameters relative to baseline; not active parameters per token.
Equation: [(m.parameters / p0).to('').magnitude for m in sparse_models].
bandwidth_growth
engine.scenario_types.ParameterBandwidthGrowthScenario.bandwidth_growth
Largest dimensionless bandwidth index in the selected accelerator series.
Equation: max(bandwidth_index).
parameter_growth
engine.scenario_types.ParameterBandwidthGrowthScenario.parameter_growth
Largest dense-model parameter index; sparse models are excluded.
Equation: max(dense_index).