The Multi-Agent Workflow GPU Requirements Estimator approximates accelerator capacity for workflows that run several model-backed agents. It converts workflow traffic into aggregate generation demand, then accounts for parallel-agent peaks, target utilization, and reserve capacity.
The estimator is best used with measured per-GPU throughput from the intended model and serving stack. It helps expose a key multi-agent issue: average token demand may look manageable while synchronized parallel branches create a much higher instantaneous GPU requirement.
1. Enter workflow traffic
Use expected completed multi-agent workflows per minute.
2. Describe agent activity
Enter model-backed agents, calls per agent, and generated tokens per call.
3. Estimate the parallel peak
Choose the share of calls likely to overlap across agents.
4. Use measured GPU throughput
Enter effective output throughput for one GPU at the intended configuration.
5. Add headroom
Set utilization and reserve, then review average and peak-adjusted demand.
The rounded count is a peak-aware throughput estimate for the whole multi-agent workload.
The parallel adjustment is a planning heuristic. Replace it with measured peak concurrency from traces when available.
Given: 10 workflows/minute, 4 agents, 2 calls per agent, 500 tokens per call, 50% parallel share, 250 tokens/second per GPU, 75% utilization, and 20% reserve.
Calculation: Average rate = 10 × 4 × 2 × 500 ÷ 60 = 666.67 tokens/second. Peak factor = 1 + 0.50 × (4 − 1) = 2.5. Peak rate = 1,666.67 tokens/second. Planned GPUs = 1,666.67 ÷ (250 × 0.75) × 1.20 = 10.67.
Result: Round up to 11 GPUs.
Simultaneous agent calls concentrate token generation into a shorter interval and raise peak compute demand.
Calculate each model pool separately when agents use different models or GPU types.
Only if their load is converted into an equivalent measured GPU throughput demand; separate pools are usually clearer.
Set parallel share near 1% or model the exact time profile separately; the heuristic is designed to capture overlapping branches.
Yes. Reserve can cover failures, rolling updates, burst traffic, and forecast error.