OmniBus™ PCIe 5.0 · SFF-8614 · Gen 5.0 Host Adapter

One fabric, every node

32.0 GT/s per lane. Sub-microsecond host-to-host latency. The OmniBus™ PCIe 5.0 NTB Adapter turns cabled PCIe into a clustering fabric — from a single hot-add to a 60-node switch — so servers, GPUs, and storage address each other as local resources and AI runs against data in place rather than across a network.

Host NTB adapter · passive heatsink

500nsApplication-to-application latency
32 GT/sPer lane (PCIe 5.0)
60Nodes on one switched fabric
01

OmniBus™ features

PCIe 5.0 performance, cabled like a network, addressed like memory.

PCI Express 5.0 — 32.0 GT/s per laneMicrochip Switchtec® PCIe 5.0 PFXLink compliant with PCIe 1.0–5.0Selectable host clock isolationAutomatic CFC / SSC supportQuad SFF-8614 x4 connectorsPassive copper cable supportRDMA via PIO and DMAEEPROM multi-system configsLink status LEDsLow profile, half length
02

Host-to-host latency

Cabled PCIe that behaves like a backplane, not a network.

500ns application-to-application latency, with chip latency near 100ns.

Fast enough that clustering and hot-add applications treat remote PCIe devices as local ones — application developers get PCIe 5.0 bandwidth without the overhead of a conventional network fabric.

PIO — optimized for small packetsDMA — large transfers, minimal CPUNon-transparent bridging
03

Configurations

The same adapter scales from one cable to a 60-node matrix.

Every topology below runs on the same OmniBus™ hardware and software suite — no separate architecture for scale.

Single-node · transparent hot add · 1 hostDirect cable · NTB network · 2 nodesMesh cable · no external switch · 3–5 nodesSwitched fabric · Gen5 24-port switch · 6–60 nodes
04

OmniBus™ software

A full clustering suite, licensed with every card.

SISCI API — shared-memory programming environment for PCIe.

Omni Sockets — Berkeley Sockets–compatible accelerator, standard TCP/IP drivers.

SmartIO — access to remote PCIe devices as if they were local.

Board Management — configuration and health visibility across the fabric.

Compatible with Linux and Windows. PIO and DMA transfer schemes are handled automatically — PIO minimizes latency on small packets, DMA moves large packets with minimal CPU utilization.

Specifications

The numbers that matter for a deployment decision.

PCI Express
Base Spec 5.0 · Cabling Spec 5.0
Max link speeds
1,024 Gb/s (32-line) · 512 Gb/s (16-line)
App-to-app latency
~500–600ns (system dependent)
Boards per system
Up to 4 (system dependent)
Usage mode
Non-transparent bridging (NTB)
Active components
Microchip Switchtec® PCIe 5.0 PFX
Cable connections
SFF-8614, quad x4, copper up to 3m
Topologies
2-node direct · 3–5 node direct · multi-node switch
Operating systems
Windows, Linux
Form factor
Low profile, half length (167.65 × 68.90mm)
Operating temp
0°C – 45°C
Typical power
12V: 2.5A (copper cable)

The cable is the cluster.

OmniBus™ turns PCIe 5.0 into a composable fabric — hot-add today, a 60-node matrix tomorrow, without a second adapter or a second architecture.

OmniBus™ PCIe 5.0 · SFF-8614 NTB Adapter

Go deeper

Explore everything on Connectivity Platform