GPU#

Bases: HasConsumptionAndEmbodiedImpacts

Subclass of Component representing a GPU.

This class uses the same estimation as MLCA, described in

Clément Morand, Anne-Laure Ligozat, Aurélie Névéol. MLCA: a tool for Machine Learning Life Cycle Assessment. 2024 10th International Conference on ICT for Sustainability (ICT4S), Jun 2024, Stockholm, Sweden. pp.227-238, 10.1109/ICT4S64576.2024.00031. hal-04643414"

See the original implementation made by the authors : https://github.com/blubrom/MLCA/blob/d2d8a42a030ea6259959b7d13a2b8637c8501044/boaviztapi/model/components/component.py#L568C7-L568C19

name

The name of a GPU model present in the database. Can be an instance of ExactName or the return value of find_close_cpu_model_name() for fuzzy matching the name. Defaults to None.

Type:: Name | None, optional

tdp

The Thermal Design Power of the GPU, used to determine the GPU’s instantaneous power consumption. Defaults to None.

Type:: ReplicableValue | None, optional

die_size

The die surface size of the GPU. Defaults to None.

Type:: ReplicableValue | None, optional

ram_size

size of the GPU’s embedded RAM. Defaults to None.

Type:: ReplicableValue | None, optional

family

Family of the GPU. Defaults to None.

Type:: str | None, optional

release_range

A range in which the GPU was released. Defaults to the last ten years.

Type:: DateRange, optional

Methods Summary

Methods Documentation

_estimate_die_size() → ReplicableValue

Estimate the die size of the GPU.

If the user provided name, we try to find in the database if the die size is known for this model. If the user provided family, we average the die sizes of GPUs in the same family. Otherwise, we return the average die size of all GPUs in the database.

Returns:: The estimated die size of the GPU.
Return type:: ReplicableValue

_get_gpu_within_date_range() → DataFrame

Return a sub-dataframe of gpu_specs containing only the rows within release_range.

Returns:: A sub-dataframe of gpu_specs containing only the rows within release_range.
Return type:: pandas.DataFrame

_get_gpus_matching_by_name() → DataFrame

Return a sub-dataframe of gpu_specs containing only the row matching name.

Return an empty dataframe with the same columns as gpu_specs if name is None.

Raises:: ValueError – name should be an instance of the abstract class Name or None. Therefore, it can be an ExactName, a FuzzymatchSingleResult, a FuzzymatchMultipleResult, or None.
Returns:: A sub-dataframe of gpu_specs containing only the row matching name.
Return type:: pandas.DataFrame

_ram_size() → ReplicableValue | None

Returns the ram size based on the name and the database. If the RAM size isn’t in database or name is not knowm, returns None.

Returns:: the ram size based on the name and the database. If the RAM size isn’t in database or name is not knowm, returns None.
Return type:: ReplicableValue | None

estimate_electric_power(workload: ReplicableValue | int | float | None = None) → ReplicableValue

Estimate the electric power of the CPU.

If known, we return the TDP. If it is not known, we average the TDP of the GPUs within the release range. If no GPUs are in the release range, we average all TDPs of GPUs in the database.

Parameters:: workload (ReplicableValue, optional) – The workload of the CPU, between 0 and 100, as a ReplicableValue. The unit should be dimensionless. Defaults to the value defined in the config file under default_values_gpu > workload.
Returns:: The estimated instant consumption of the CPU.
Return type:: ReplicableValue

estimate_embodied_impacts() → Impacts

Estimate the embodied impacts of the GPU.

GPU embodied impacts are estimated using: self._die_size_factor * self.die_size + self._base_impact, where: - _die_size_factor is the impact of 1 mm² of die. It is defined in the config file under default_values_gpu > die_size_factor. - die_size is the surface of the die in mm², either provided by the user or estimated in _estimate_die_size(). - base_impact is the impact of everything that isn’t the die in a GPU, and is defined in the config file under default_values_gpu > base_impact.

To this, we add the RAM impact, estimated using the RAM class.

Returns:

The embodied impacts of the GPU.

Return type:

Impact

estimate_idle_power() → ReplicableValue | None: Returns the minimum consumption of the component, the consumption it has when no job are running

estimate_peak_power() → ReplicableValue | None: Returns the maximum consumption of the component