GPU "Napkin Math"

How fast is energy consumption in LLMs growing?

kWh

1 kWh = 1 Block

1 Laptop (Daily Usage)(~0.2 kWh)

Single GPU (H100, per hour)(~0.7 kWh)

GPT-1 Training(100 kWh)

MWh

1 MWh = 1,000 kWh

GPT-2 Training(1 MWh)

1 US Home (Annual)(~10 MWh)

Geohotz Tinybox Pro(~70 MWh)

GWh

1 GWh = 1,000 MWh

Stanford Natural Language Computing (64 H100, annual) (~392 MWh)

GPT-3 Training(1.287 GWh)

GPT-4 Training (Estimate)(10 GWh)

Small City (Annual)(~20 GWh)

TWh

1 TWh = 1,000 GWh

Meta Cluster (350k H100s, Annual)(~2.15 TWh)

Nuclear Power Plant (Annual)(~7.88 TWh)

OpenAI 2025 Cluster (Estimate)(~100 TWh)

PWh

1 PWh = 1,000 TWh

US Power Grid (Annual)(~4 PWh)

World Power Grid (Annual)(~30 PWh)

kWh

MWh

GWh

TWh

PWh

Inspired by github.com/chubin/late.nz [MIT License]

And from "Jeff Dean's latency numbers"