Estimate the accelerator fleet needed for a token-based AI document processing workload. The calculation converts daily document and page volume into total model tokens, divides that demand by measured per-GPU throughput and the available processing window, then adds explicit capacity headroom. Use benchmark throughput from your own model, precision, batch size, and GPU type. OCR running on CPUs or a separate service should not be included unless its work is represented in the token load.
Daily tokens = Documents × [(Pages × Tokens/page) + Tokens/document]
GPUs = ceil{[Daily tokens ÷ (Tokens/sec/GPU × 3,600 × Hours × Utilization)] × (1 + Headroom)}
The main result is an integer provisioning estimate. Base GPU load shows the fractional compute demand before headroom and rounding.
Memory capacity, context length, model replicas, multi-GPU tensor parallelism, and latency targets may require more GPUs than this throughput-only estimate.
50,000 documents × [(6 × 800) + 400] = 260,000,000 tokens/day. At 1,200 tokens/second, 20 hours, and 70% utilization, base load is 4.30 GPUs. Adding 20% headroom and rounding up gives 6 GPUs.
Use measured throughput from the same model, precision, context mix, and batching configuration planned for production.
No. The formula sizes throughput only; memory limits and multi-GPU model placement must be checked separately.
GPU count is discrete, so a small increase above an integer boundary requires one additional whole GPU.
Include the text tokens sent into the model, but do not convert CPU OCR compute time into tokens for this GPU calculation.
Add enough separately provisioned replicas to satisfy your failure-domain policy if the headroom field does not explicitly cover that requirement.
| Input | Purpose | Unit |
|---|---|---|
| Token workload | Daily model demand | tokens/day |
| GPU rate | Measured throughput | tokens/second/GPU |
| Utilization | Effective busy fraction | percent |
| Headroom | Extra capacity before rounding | percent |