Scenarios Zoo

Workload Statistics, Scale Profiles, and Comparison Anchors

The Scenarios Zoo holds reusable narrative and teaching anchors: real-world workload statistics, energy and emissions comparison baselines, mobile power envelopes, phone batteries, clinical-imaging workflow anchors, storage running examples, and distributed-training scale profiles. These are not hardware specs and not literature categories; provenance records where the number came from, while the registry path records what kind of fact it is.

Use Scenarios.* for reusable problem settings that appear across examples. Use Hardware.* for device specifications, Systems.* for composed physical systems such as nodes/racks/clusters/storage paths, Infrastructure.* for facility, pricing, grid, and capacity facts, and Ops.* for operational policies or overhead profiles. A one-off hypothetical knob can stay local in a LEGO cell when it defines the local problem rather than a shared fact.

Workload scale anchors

Entry Value Description
Gmail emails per day 121000000000.0 Approximate daily Gmail volume.
Google searches per day 13700000000.0 Lower bound from Google’s disclosure of more than 5 trillion searches a year (2025).
Waymo sensor data rate (high) 19.0 TB/h Upper-bound AV sensor data generation rate.
Waymo sensor data rate (low) 1.0 TB/h Lower-bound AV sensor data generation rate.

Clinical imaging anchors

Entry Value Description
Retinal screening image size 40000000.0 Reference size for one retinal screening photograph in the rural-clinic workflow.

Energy comparison anchors

Entry Value Description
Boiling a cup (~250 mL) of water 100000 J Approximate energy to bring one cup (~250 mL) of water to a boil.
Smartphone full charge 40000 J Approximate energy to fully charge a smartphone battery.
US household annual electricity 10.7 MWh Rounded annual electricity use baseline for one average US household-year.

Emissions comparison anchors

Entry Value Description
Transatlantic round-trip passenger CO₂e 1000.0 One economy passenger, New York to London and return (kg CO₂e).

Training scale profiles

Entry Value Description
Scaling efficiency (1024 GPUs) 0.83 Modeled for the 175B reference job; pipeline fill and drain become visible (~6% of step).
Scaling efficiency (256 GPUs) 0.87 Modeled for the 175B reference job at a fixed GPT-3 batch; tensor-parallel communication is the main loss.
Scaling efficiency (32 GPUs) 0.9 Book convention for small jobs; the 175B reference job needs 64 GPUs per TP8xPP8 replica, so this tier is not modeled.
Scaling efficiency (8192 GPUs) 0.56 Modeled for the 175B reference job at a fixed GPT-3 batch; pipeline bubble dominates (~33% of step). Production runs with larger batches and interleaved pipelines retain more (MegaScale).
Per-accelerator compute efficiency 0.45 Fraction of peak each accelerator achieves on its local compute before communication losses.

Storage training corpus

Entry Value Description
Storage running-example checkpoint bytes per parameter 112.0 Resumable mixed-precision Adam checkpoint footprint: FP16 weights, FP32 master weights, and two FP32 moments; gradients excluded.
Storage running-example checkpoint interval 10 min Reference checkpoint interval for checkpoint-count examples.
Storage running-example compressed source corpus 24000000000000.0 Compressed source corpus size for the storage running example.
Storage running-example training step time 200 ms Reference per-step compute time the input pipeline must keep fed.
Storage running-example token ID width 32.0 Serialized token-ID width for the tokenized corpus.
Storage running-example tokenized shard size 24000000000.0 Reference size of one tokenized training shard file.
Storage running-example tokenized corpus 48000000000000.0 Derived serialized token-ID corpus size for one epoch.
Storage running-example training tokens 1500000000000.0 Reference token count for the 175B-model storage running example.
Storage running-example training window 30 d Reference training-window duration for checkpoint-count examples.

Mobile power envelopes

Entry Value Description
Background adaptation power budget (high) 1000 mW Upper bound for background on-device adaptation power.
Background adaptation power budget (low) 500 mW Lower bound for background on-device adaptation power.
Mobile ML burst power (high) 10 W Upper bound for brief smartphone ML burst processing.
Mobile ML burst power (low) 5 W Lower bound for brief smartphone ML burst processing.
Mobile ML sustained power (high) 3 W Upper bound for sustained smartphone ML processing.
Mobile ML sustained power (low) 2 W Lower bound for sustained smartphone ML processing.
Mobile NPU peak power 4 W Higher-bound mobile inference power envelope.
Mobile NPU typical power 3 W Typical sustained power for on-device mobile inference.
Object-detector power 2 W Reference power draw of an always-on object-detection workload.

Phone battery anchors

Entry Value Description
Phone battery capacity 3000 mAh Typical flagship smartphone battery charge capacity.
Phone battery energy (capacity x voltage) 39960.0 J Battery energy derived from capacity x voltage.
Flagship phone battery energy 15 h·W Modern flagship smartphone battery pack energy rating.
Phone battery voltage 3.7 V Nominal Li-ion cell voltage.

Python Access

import mlsysim

eff_1024 = mlsysim.ReferenceStats.TrainingScaleProfiles.Eff1024Gpu
retinal_photo = mlsysim.ReferenceStats.ClinicalImaging.RetinalPhotoSize
flight_kg = mlsysim.ReferenceStats.EmissionsAnchors.TransatlanticRoundTripCo2Kg
household_year = mlsysim.ReferenceStats.EnergyAnchors.USHouseholdAnnualElectricity
corpus_tokens = mlsysim.ReferenceStats.StorageTrainingCorpus.TrainingTokens
checkpoint_interval = mlsysim.ReferenceStats.StorageTrainingCorpus.CheckpointInterval
Back to top