Fabric calculator

Put in how many NIC ports you need to connect (for a GPU cluster, usually GPUs times NICs per GPU), the port speed, and the switch chip. The calculator uses textbook leaf-spine and fat-tree maths with no oversubscription. It runs in your browser; nothing you type is sent anywhere.

How it works

A switch chip's capacity divided by the port speed gives its radix, the number of ports. A 102.4T chip has 64 ports of 1.6T; a 204.8T chip has 128.

• One tier: a single switch, up to one endpoint per port.
• Two tiers (leaf and spine): each leaf uses half its ports for endpoints and half to reach the spines, so the fabric tops out at radix times half the radix. That's 2,048 endpoints at 64 ports, 8,192 at 128.
• Three tiers (a fat tree): up to radix cubed divided by four. That's 65,536 endpoints at 64 ports.

Every link counts as two pluggable ends, one at each end, including the NIC side. In a real build some short links use copper cables (DAC or AEC) instead of optics, and many fabrics oversubscribe the upper tiers, so treat the counts as the non-blocking upper bound.

The switch chips

• 51.2T (256 lanes of 200G). For example Cisco Silicon One G302, a top-of-rack chip due in Q4 2026.
• 102.4T (512 lanes of 200G). Shipping, for example Cisco Silicon One G300.
• 204.8T (512 lanes of 400G). The 400G-per-lane generation USFP is built for. No such chip announced yet.

The 204.8T row is there to show the effect of 400G lanes. No 400G-per-lane switch chip has been announced, so its port count is the arithmetic of 512 lanes at 400G, not a product.

Why this matters for USFP

Doubling the lane rate doubles the radix at the same port speed, and a bigger radix keeps more of a cluster in fewer tiers. That is the argument Cisco and NVIDIA make for USFP: connect 1.6T endpoints without adding tiers. The module is what lets a 400G-per-lane chip get all those ports onto a front panel.

• USFP vs OSFP

  Same footprint, four times the bandwidth, a new connector. Why your OSFP cages won't take it.

• USFP vs XPO

  Air-friendly 6.4T against a liquid-cooled 12.8T. Two camps, two bets on the next switch chip.

• USFP vs co-packaged optics

  Keep swappable modules, or put the optics in the switch package for the lowest power.

• Every form factor

  OSFP to co-packaged optics in one table: lanes, capacity, heat and who backs each.