llm-resource-consumption-quantifiable-across-three-dimensions

IN derived (depth 1)

Created 2026-06-21T13:10:22+00:00 · Reviewed 2026-06-21T14:41:08+00:00

LLM resource consumption is precisely quantifiable across three independent dimensions — memory (2 bytes per float16 parameter), compute (6 FLOPs per parameter per token for training, 1-2 for inference), and energy (0.002-0.05 Wh per prompt depending on task type) — enabling principled cost modeling for deployment decisions.

Summary

Running a large language model costs resources in three predictable ways — how much memory to store it, how much compute to run it, and how much electricity each query burns — and because each has a known per-unit rate, you can calculate the exact cost of a deployment before committing hardware. This turns infrastructure planning from guesswork into arithmetic: pick a model size, multiply by the per-parameter and per-token rates, and you get a defensible budget for memory, compute, and energy.

Justifications

SL — Three independent resource metrics provide concrete accounting for the full deployment cost structure

Antecedents (all must be IN):

IN float16-parameter-size-2bytes — One float16 parameter requires 2 bytes of storage, so 1 billion float16 parameters require approximately 2 GB.
IN training-cost-6-flops-per-param-per-token — LLM training costs approximately 6 FLOPs per parameter per token, while inference costs 1-2 FLOPs per parameter per token.
IN energy-per-prompt-classification-vs-generation-vs-image — Energy cost per prompt varies by task type: classification ~0.002-0.007 Wh, text generation ~0.05 Wh, image generation ~2.91 Wh (up to 11.49 Wh for least efficient models).