> ## Documentation Index
> Fetch the complete documentation index at: https://docs.protectorplus.cloudsine.tech/llms.txt
> Use this file to discover all available pages before exploring further.

# Performance

> Latency and throughput on the recommended hardware specification.

Performance figures are obtained on the CloudsineAI recommended hardware specification. Latency is measured under no-load conditions; throughput is measured under the stated maximum concurrent-user load with the listed prompt sizing.

## Recommended hardware

| Specification     | Value              |
| ----------------- | ------------------ |
| AWS Instance Type | `g6e.2xlarge`      |
| vCPU              | 8                  |
| GPU               | NVIDIA L40S, 48 GB |
| Memory            | 64 GB              |
| Storage           | 450 GB             |

## Test profile

| Test profile                 | Prompt size                       | Latency SLO              |
| ---------------------------- | --------------------------------- | ------------------------ |
| Input-guardrail performance  | 115 tokens (\~614 characters)     | ≤1.5 seconds per message |
| Output-guardrail performance | 1,150 tokens (\~6,140 characters) | ≤1.5 seconds per message |

Per-guardrail latency and throughput tables — including no-load per-message latency, maximum concurrent-user counts at the 1.5-second SLO, and the supporting CSV exports — are available on request as part of the full UAT performance report.

## Sample latencies

Indicative figures from a recent benchmark run (LLM-only configuration against a public prompt-injection corpus):

| Metric       | Value             |
| ------------ | ----------------- |
| Mean latency | \~1.3 seconds     |
| P95 latency  | \~1.5–1.6 seconds |
| Precision    | \~97%             |

Exact values vary by configuration (guardrail mix, vector sensitivity) and by dataset.