Model hardware calculator

From model to GPU, priced.

Choose a model and estimate its memory needs. Compare published GPU hardware and rental prices using your precision, context and deployment settings.

Selected Model

bert-base-chinese

google-bert/bert-base-chinese

Change model View model details

Parameter count: 102,882,442
Architecture: bert

Choose your workload

Memory and rental estimate

Model weight precision Nominal storage calculation; selecting a precision does not convert or quantize model files.

KV-cache precision assumption Some runtimes may not support each cache format for a given model.

Context tokens A repository position limit is not automatically treated as your serving context.

Concurrent requests Used only in the KV-cache calculation when architecture fields are known.

Deployment configuration Multi-GPU and CPU-offload results stay conditional on provider-published hardware and framework details.

Rental hours per day

Rental days per month

Provider region Only regions present on a current GPU offer are listed.

Search published GPU offers

Sort offers Prices in different currencies are not converted or ranked against each other.

Method and limits

How memory and offer checks work

Weights: total reported parameter count × the selected nominal bits per parameter ÷ 8. This is a simplified storage estimate. INT8 and INT4 overhead and unquantized components are not inferred when the repository does not report them. For mixture-of-experts models, all reported parameters remain in the estimate.

KV cache: when published configuration fields support it, layers × requested context × concurrency × KV heads × explicit head dimension × 2 for key and value × selected cache bytes. Missing dimensions leave KV cache unknown; the tool does not infer head dimension from architecture names.

Hosting: only current GPU offerings appear. A lower-bound memory estimate is only a first check. Runtime and system overhead can change the result. We show “meets documented minimum” only when the documented hardware and deployment setup establish it. Unknown provider charges stay unknown.

Accelerate's model estimator builds an empty model from configuration metadata and estimates loading memory. Its documentation explicitly separates model-loading memory from inference memory; this tool follows that boundary and does not treat its own subtotal as total runtime requirements.