Skip to content

GPU

Characteristics

Name Unit Default value Description Example
units None 1 Number of GPUs (see gpus_per_card) 2
usage None See Usage See usage ..
name None None Full commercial name of the GPU NVIDIA H100 SXM 80GB
weight kg 1.69 Total mass of the GPU card 1.69
heatsink_weight kg 0.90 Mass of the heatsink 0.90
casing_weight kg 0.79 Mass of the casing / shroud 0.79
gpu_surface mm² 814 (effective) Effective GPU die area including wafer losses 2810.4
vram GB 80 Total VRAM capacity of the card 80
vram_dies None 6 Number of VRAM dies 6
vram_surface mm² computed Total effective VRAM die area including wafer losses 744.0
pwb_surface cm² 296 PCB surface area 296.37
pwb_weight kg None Mass of the PCB 0.2
tdp W 700;70;1100 Board power of one GPU card (TDP) 700
gpus_per_card None 1 Number of GPUs on one card, e.g. 2 for a Tesla M60 2
transport_boat km 19000 Distribution distance by boat 19000
transport_truck km 1000 Distribution distance by truck 1000
transport_plane km 0 Distribution distance by plane 0

Complete

The following completion strategies can be used.

Completion from GPU name

If a GPU name is provided, the following attributes can be retrieved via fuzzy matching on the GPU name repository:

  • vram, vram_dies
  • gpu_surface (already the effective area — used as-is)
  • pwb_surface
  • weight, heatsink_weight, casing_weight
  • transport_boat, transport_truck, transport_plane
  • tdp, gpus_per_card

Warning

The GPU name repository is not exhaustive, and the fuzzy match may return a different GPU than the one specified if the name is ambiguous or misspelled. The matching threshold is configurable via BOAVIZTA_GPU_NAME_FUZZYMATCH_THRESHOLD in the config. You can contribute new GPU models by following the contributing guide.

Completion of vram_surface

If vram_surface is not set, it is computed from vram and vram_dies:

\[ \text{die\_area} = \frac{\text{vram} \times 15.5}{\text{vram\_dies}} \]
\[ \text{vram\_surface} = \text{effective\_area}(\text{die\_area}) \times \text{vram\_dies} \]

where \(\text{effective\_area}\) accounts for wafer losses and die yield using the standard circular wafer model.

Completion of gpu_surface

If gpu_surface is not set and not completed from name, the archetype default value (already expressed as effective area) is used.

Embedded impacts

Impact criteria

Criteria Implemented
gwp yes
adp yes
pe yes
gwppb no
gwppf no
gwpplu no
ir yes
lu yes
odp yes
pm yes
pocp yes
wu yes
mips yes
adpe no
adpf yes
ap yes
ctue yes
ctuh_c no
ctuh_nc no
epf no
epm no
ept yes
fw no
fe no

Impact formula

The total embedded impact of one GPU card is the sum of contributions from each sub-component:

\[ \text{GPU}_\text{embedded}^\text{criteria} = \text{impact}^\text{casing} + \text{impact}^\text{heatsink} + \text{impact}^\text{pwb} + \text{impact}^\text{gpu\_die} + \text{impact}^\text{vram} + \text{impact}^\text{upstream\_transport} + \text{impact}^\text{boat} + \text{impact}^\text{truck} + \text{impact}^\text{plane} + \text{impact}^\text{end\_of\_life} \]

Each sub-component impact uses a dedicated impact factor from the factor database:

  • Casing / heatsink: factor × mass (kg)
  • PWB: factor × pcb surface (cm²)
  • GPU die: factor × effective die area (mm²)
  • VRAM: factor × effective total VRAM die area (mm²)
  • Upstream transport: factor × total card weight (kg)
  • Distribution transport (boat / truck / plane): factor × distance (km) × total card weight (kg)
  • End of life: factor × total card weight (kg)

Info

The characteristics above describe a whole card. The embedded impact of one card is multiplied by units / gpus_per_card, the number of cards: providers count the GPUs of a dual-GPU card such as the Tesla M60 separately, so units is a number of GPUs.

Usage impacts

Power consumption is based on the usage methodology. The GPU's workload and use time ratio are taken from the usage parameters.

avg_power is the average power of one GPU. When it is not given, it is modelled from the tdp, linearly from an idle power of 10% of the TDP to the full TDP. The idle share is in line with the minimum and maximum powers of the Cloud Carbon Footprint GPU methodology:

\[ \text{P}_{\text{GPU}}(\text{workload}) = \frac{\text{tdp}}{\text{gpus_per_card}} \times (0.1 + 0.9 \times \frac{\text{workload}}{100}) \]

The workload is the GPU's own time_workload. It does not follow the time_workload of a server or cloud instance, which describes its CPU. By default, it is assumed to be 65%. This is an assumption, not a measurement: it sits in the 60 to 70% utilization of AI datacentre GPUs that a Google architect is reported to have given in Tom's Hardware, in an article on GPU service life. When the actual load is known, set gpu_time_workload in the usage of the server or cloud instance, or time_workload in the GPU's own usage. Total power is \(\text{P}_{\text{GPU}} \times \text{units}\).