EXPERIMENTAL LAB · EMBEDDED

LLM Cluster

Plan distributed training and inference, then open the complete same-origin simulator.

FIRST-PARTY LLM CLUSTER PLANNER

Design the topology, then inspect the tradeoffs.

Open the complete interactive workspace for the full simulator, Learn mode, and Play mode. Use the compact first-party projection below for a fast comparison.

CONFIGURATION

Workload and hardware

ModelGPUPrecisionGPU countTensor parallelPipeline parallelBatch sizeSequence lengthOutput tokens

PROJECTION

Capacity and performance

✓ Fits estimated GPU memory22.9 / 72 GB per GPU

Throughput2,186.5 tok/s

Model FLOP utilization54.4%

Communication share2.8%

Data parallel1×

Hourly cluster cost$31.2

Cost / 1M tokens$3.96

Estimated generation time1.87 s

Primary bottleneckMemory bandwidth

Recommendations

Increase batch reuse or choose a GPU with more memory bandwidth before adding replicas.

Estimator assumptions and limits

Analytical estimate only; no model weights or GPU workload runs in this browser.
Published peak hardware figures are discounted by utilization and topology penalties.
Cloud prices are editable-release assumptions and exclude storage, networking, reservations, support, and tax.

Compact Frontier Rehearsal estimator v1. Hardware and price inputs are planning assumptions, not vendor quotes or benchmark guarantees. Validate shortlisted configurations against the exact framework, kernel, topology, provider, and workload before spending money.

SESSION BOUNDARY

Local to this tab.

Runs stay in your browser and end when you leave the page. No account, saved progress, cloud resource, or private practice data is attached to this lab.

Preparing this page…

Your current navigation and account state stay available while the next page loads.