Vendor-neutral design tools for AI infrastructure networking

What does the network for N GPUs actually look like, and what do you need to buy? Interactive tools with the math validated against published vendor reference architectures and peer-reviewed research — sources cited, revisions pinned, divergences documented.

Before you size anything: what kind of cluster is this?

The single biggest AI network design decision isn't a switch model — it's your serving architecture, and no vendor document tells you the whole story.

Training clusters need a dedicated, lossless east-west backend fabric. Every published reference architecture agrees.

Inference clusters split two ways, and the published guidance diverges:

• Data-parallel serving (one model instance per GPU): validated designs from multiple vendors show the backend fabric can be omitted entirely — the frontend carries everything. Caveat: a cluster built this way can't take on training or hybrid workloads without disruptive retrofit.
• Disaggregated prefill/decode serving (the default in modern production stacks): an east-west RDMA fabric is required — not for gradients, but for KV-cache transfer between prefill and decode pools. (The published research assumes this fabric and shows the transfer cost stays small when it's there — the fabric is what makes disaggregation work, not an optional add-on.)

This tool asks the question before doing the math, and sizes accordingly. We're vendor-neutral: every calculation is validated against published reference designs from NVIDIA, Cisco, Juniper, and Asterfusion, plus peer-reviewed research — with hardware data drawn from published NVIDIA, AMD, Arista, Broadcom, and HPE Juniper documentation. Sources cited, revisions pinned, divergences documented.

We sell nothing and favor no one.

Or jump straight into a worked example:a spineless 1,024-GPU cluster8,192 GPUs on Tomahawk 6disaggregated inference