What's a "Cluster"?

A cluster is simply a group of separate computers connected together so they can share the work and act, from the outside, like one much bigger machine.

The Honest Version: Real Clustering vs. Stacking Desktop Cards

That one sentence is true, but it hides a huge gap in what "connected together" actually means in practice. There are two very different things people call a "cluster," and only one of them is what real AI companies run in production.

✅ Real Datacenter Clustering

Inside one server, GPUs talk to each other over NVLink — a dedicated connection built for this, moving data at roughly 900 GB/s, hundreds of times faster than a regular cable. Between servers, an InfiniBand network moves data at 400+ Gb/s with almost no delay.

On top of that hardware, real clusters run coordination software (like NCCL, Slurm, or Kubernetes) that splits a huge model's memory precisely across every GPU and keeps all of them working on the same problem in lockstep. This is what lets 16 or 100+ GPUs genuinely behave like one giant machine.

⚠️ Just Stacking Desktop Cards

Putting several RTX 4090s in one home PC connects them over regular PCIe slots — about 32–64 GB/s, roughly 15–25x slower than NVLink, and consumer GPUs since the 2020s don't support direct GPU-to-GPU links at all.

Connecting several separate desktop towers over a home or office network (1–10 Gb/s Ethernet) is slower still, and lacks any of the coordination software that keeps a real cluster in sync. It can work for small hobby experiments split across a couple of GPUs in one box — it is not how any company serves AI to real customers.

Why It Matters When You're Buying Hardware

If someone offers to sell you a "cluster" that's really just several gaming PCs on a network switch, you are not getting production-grade AI infrastructure — you're getting a science project. Our Mid-Size Company build uses real InfiniBand networking for exactly this reason.