CONFLOW POWER GROUP

CONFLOWPOWER.COM

Edge vs hyperscale: latency calculator

Compare real-time inference latency between an iLamp distributed edge fabric and a centralised hyperscale data centre. Adjust the workload and client distance to see how round-trip times change for each path.

200 km
500 m

iLamp edge fabric

— ms
single radio hop

Hyperscale data centre

— ms
via public internet

Methodology

The model assumes fibre-effective light speed of 200,000 km/s for the data-centre backbone (accounting for refractive index and routing overhead), a 2 ms round-trip radio hop for the edge path, and 8 ms of ISP and access-network overhead on the hyperscale path before traffic reaches the long-haul backbone. Inference time is held constant across both paths, since the same NVIDIA Jetson AGX Orin silicon (275 TOPS) is the proxy for compute — isolating the variable that actually matters: where the work physically happens.

Vision time reflects single-frame YOLO/ResNet-class inference. Speech time reflects streaming ASR per phrase. LLM time reflects per-token generation on a 7-billion-parameter quantised model. All figures are indicative round-trip totals from request emission to response receipt.