AcceleratorPodScenario

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 pod envelope of pod.chips copies of accelerator: aggregate peak and memory.

Notes

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

Assumptions: The pod envelope contains identical chips; capacities and peaks add arithmetically.

Excludes: Aggregate local memory is not a unified address space, and aggregate peak does not imply achieved scaling.

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.TPUv4Pod4096
>>> str(case.peak_flops)
'1126400.0 TFLOPs/s'

Attributes

Name	Description
accelerator
chip_memory_total	Sum of local chip memory capacities. Equation: pod.chips * | | [peak_flops](#mlsysim.engine.scenario_types.AcceleratorPodScenario.peak_flops) | Sum of per-chip arithmetic peaks in the pod envelope. Equation:pod.chips *
pod

pod

Type: object.

Declared default: required.

accelerator

Type: HardwareNode.

Declared default: required.

peak_flops

engine.scenario_types.AcceleratorPodScenario.peak_flops

Sum of per-chip arithmetic peaks in the pod envelope. Equation: pod.chips * accelerator.compute.peak_flops. Units: FLOP/s; perfect parallel utilization is not implied.

chip_memory_total

engine.scenario_types.AcceleratorPodScenario.chip_memory_total

Sum of local chip memory capacities. Equation: pod.chips * accelerator.memory.capacity. Units: bytes; this does not make the memory a single shared address space.