Product Brief

Converge Servers and IO controllers with PCIe

n

Create cost-effective high-availability hyperscale systems

by enabling communication between in-rack hosts and

endpoints using PCIe

n

Simplify connectivity while providing the highest PCIe

switching performance available for data center servers,

storage, and networks

n

Reduce latency, system complexity, and power consumption

by up to 50% in data-intensive environments

n

Take advantage of industry-first features for most demanding

hyper-converged, NVMe and rack scale systems

Avago PEX9700 switches allow customers to build high performance,

low latency, scalable, cost-effective PCI Express-based fabrics. The

switches enable I/O sharing with standard SR-IOV or multifunction capability, allowing multiple hosts

or Nodes to reside on a single PCIe-based network. Hosts communicate through Ethernet-like DMA

(NIC DMA) with other hosts and end-points using application software. Hosts may also communicate

using Tunneled Window Connection (TWC), a special low latency host-to-host communication

capability for short packets.

Shared I/O Using Standards

PEX9700 switches allow the Virtual Functions (VFs) of SRIOV endpoints (such as an Avago

MegaRAID® SAS controller) to be shared and assigned to multiple hosts concurrently. Each host

can enumerate its assigned functions using standard BIOS and OS software and use them with

unmodified vendor-supplied drivers. The use of standard system software minimizes software

support costs.

General Purpose Host-to-Host DMA

Ethernet is used almost universally for server to server communications. PEX9700 switches

contain a virtual Ethernet NIC at each host port that allows Ethernet to be tunneled transparently

through the fabric to any and all servers that are connected. Internal Ethernet communications

using virtual Ethernet NICs and NIC DMA are complemented by the ability to share a physical

SRIOV NIC, thus providing compatibility with the vast library of applications that leverage

Ethernet communications.

Software-Defined Fabric

The switches are built on a hybrid hardware/software platform that offers high configurability and

flexibility in regards to the number of hosts, end-points, and PCIe slots. The critical pathways have

direct hardware support, enabling the fabric to offer non-blocking, line speed performance with

features such as I/O sharing and DMA. The solution is completed by management processor that

communicates with platform management via API and/or CLI. The solution offers an innovative

approach to setup and control, making use of an off-chip management CPU (mCPU) to initialize

the PEX9700 switch, configure the routing tables, handle errors, Hot-Plug events, and enable the

solution to extend the capabilities without modifying the system software.

General Features

■

n

State-of-the-art switch fabric

– Sharing I/Os among multiple hosts

– Host-to-host DMA

– Low latency TWC

■

n

Any port can be a host port or

Downstream (device) Port

■

n

Works with standard PCIe end-points

and hosts – and software, as well as

with existing application software

■

n

MSI-X support

■

n

Allows flexible fabric topologies

Key Advantages

■

n

PCI Express Switches

–12 to 97 Lanes with Integrated

on-chip SerDes

–5 to 25 Independent ports

–Designate any Port as the

Upstream Port

– Low-power SerDes (under 90 mW

per Lane)

– Device-Specific Relaxed Ordering

– Port configuration

– Dedicated management port

for mCPU

– x4, x8, or x16, depending on Port

configuration; x4 can down-train to

x1 and x2 width

– Configurable through serial EEPROM,

I2C, SMBus, and/or Host port

■

n

Standards Compliant

– PCI Express Base Specification, r3.1

(backward compatible w/ PCIe r2.0,

& r1.0a/1.1)

– PCI Power Management Spec, r1.2

■

n

High Performance

– Full line rate on all ports

– Cut-Thru packet latency of less than

150ns (x16 to x16)

– 2KB Max Payload Size

– Multicast through DMA

PEX9700 Series Switch Chips

Managed PCI Express Switches Based on ExpressFabric® Technology