ONE REQUEST / REPEATING COMPUTE & EXCHANGE

How the hardware works together

See the roles of the CPU, GPUs, memory, and network in an AI request.

Preparing the compute tray…

Drag to orbit · click to identify

Input dataIllustrative flow · not measured time or wiring

STAGE 01 / 05

How does the request get in?

A request reaches the system through its network hardware. Think of it as the entrance for information coming from elsewhere.

Request

Network adapter

Input received

Why does this matter?

The processors need input before they can produce a useful result.

WHAT MOVES?Input data

This adapter represents several real network devices. The client and service software are outside this tray.

About this example & sources

This is an authored walkthrough, not a captured execution trace. Assume a model already loaded across several GPUs. Software can overlap or repeat these stages. The tray and external endpoints are simplified; memory, rack switches, serving software, and cooling infrastructure are not drawn to scale. Reviewed October 4, 2026.

NVIDIA · GB200 system architecture ↗NVIDIA · GPU collective communication ↗NVIDIA · CUDA programming model ↗

Receive → Prepare → Calculate ↔ Exchange → Return. Example: tensor-parallel inference; exchanges repeat across layers.

FOLLOW AN AI REQUESTCompute and exchange can repeat

Power and cooling support every stage. Playback advances every 8 seconds for reading; this is not processing speed.