SS22-0B00-03 Complete Guide: Broadcom PEX89072 PCIe 5.0 Switch

Introduction

The data center is in the middle of a fundamental architectural shift. As GPU-accelerated AI training clusters grow from dozens to thousands of accelerators, and as NVMe storage demands approach terabytes per second, the PCI Express interconnect fabric that ties these components together has become one of the most performance-critical elements of the entire system. PCIe Gen 4 at 16 GT/s served the hyperscale transition well — but the leap to PCIe Gen 5.0 at 32 GT/s doubles the per-lane bandwidth once again, arriving at exactly the moment that AI infrastructure engineers are discovering that their GPU-to-storage and GPU-to-GPU communication bottlenecks have migrated entirely onto the PCIe fabric.

The SS22-0B00-03 from Broadcom is the PEX89072B0-3-DB — the production bulk-packaging version of Broadcom's PEX89072, a 72-lane, 36-port, PCIe Gen 5.0 managed switch capable of up to 4,608 Gb/s (576 GB/s) aggregate raw bandwidth in an FCBGA-1517 package. It is part of Broadcom's PEX89000 ExpressFabric platform — a family of PCIe 5.0 switches purpose-built for composable hyper-scale compute, AI/ML accelerator fabrics, and high-density NVMe storage architectures in data centers and cloud infrastructure.

The SS22-0B00-03 represents one of the highest-bandwidth single-chip PCIe switch solutions commercially available, combining the line-rate switching of PCIe 5.0 with enterprise-grade security (Hardware Root of Trust, attestation), multi-host fabric support, SR-IOV, and Broadcom's proven ExpressFabric software ecosystem — all in a component that fits the demanding thermal and power envelope of hyperscale server designs.

---

1.0 What Is the SS22-0B00-03? An Overview

The SS22-0B00-03 is Broadcom's ordering code for the PEX89072B0-3-DB — the specific configuration of the PEX89072 PCIe 5.0 switch die in production bulk packaging. Understanding the Broadcom ordering system requires decoding two separate naming conventions:

Broadcom internal ordering code (SS22-0B00-03):

• SS22 — PEX89000 series family code designating the PEX89072 die (72-lane variant)
• 0B — Package revision B (FCBGA-1517, the production package)
• 00 — Base configuration (no special customer-specific options)
• 03 — Packaging format: bulk tray (03); tape-and-reel packaging would be a different suffix

Broadcom product identifier (PEX89072B0-3-DB):

• PEX89072 — Product family: PCIe Gen 5.0, 72 lanes, 36 ports, managed switch
• B0 — Die silicon revision B0 (current production revision)
• 3 — Speed grade / configuration designator
• DB — Die in Bulk packaging (tray format)

The companion SS22-0B00-02 corresponds to PEX89072B0-2-DB — a slightly different configuration of the same PEX89072 die. Both operate as PCIe 5.0 switches; the suffix distinction reflects specific feature configurations or speed bin qualifications within Broadcom's product portfolio. For production designs, always specify the exact SS22-0B00-03 or SS22-0B00-02 ordering code to ensure the correct configuration is received.

---

2.0 Full Technical Specifications & PCIe 5.0 Architecture

2.1 Core Lane, Port, and Bandwidth Parameters

The following parameters are drawn from Broadcom's PEX89000 series product brief and PEX89072 product page:

• Product Designation: PEX89072B0-3-DB (Broadcom ordering code: SS22-0B00-03)
• PCIe Generation: PCIe Gen 5.0 (32 GT/s per lane)
• Total Lanes: 72 lanes
• Total Ports: 36 ports (each port configurable as x1, x2, x4, x8, or x16)
• Maximum Port Width: x16 per port
• Raw Bandwidth per x16 Port: 1,024 Gb/s (128 GB/s) bidirectional
• Total Aggregate Switch Bandwidth: Up to 4,608 Gb/s (576 GB/s) raw
• Switching Architecture: Cut-through, non-blocking fabric
• Latency: Single-digit nanosecond cut-through latency (PCIe fabric-level)
• Multi-Host Support: Yes — multiple CPU/host connections to shared I/O fabric
• SR-IOV Support: Yes — standard PCIe Single Root I/O Virtualization
• ExpressFabric: Yes — Broadcom's composable PCIe fabric technology
• CXL Support: Yes (CXL 1.1/2.0 compatible where supported by the connected devices)
• Hardware Root of Trust (RoT): Yes — secure boot with Chain of Trust
• Attestation: Yes — TPM-style device identity and firmware integrity verification
• Package: FCBGA (Flip-Chip Ball Grid Array), 1,517 balls
• Package Body Size: Approximately 45 mm × 45 mm (large BGA)
• Operating Temperature: 0°C to +85°C (standard commercial)
• Power Consumption: Contact Broadcom for thermal design power (TDP) specifications — PCIe 5.0 switch ICs at this lane count typically consume 30–60W depending on traffic load and configuration

---

2.2 ExpressFabric, Multi-Host, SR-IOV, and CXL Support

The PEX89072's most architecturally significant feature beyond raw PCIe 5.0 speed is its ExpressFabric capability — Broadcom's branded implementation of composable PCIe fabric infrastructure:

Base Operation (Single-Host): In standard PCIe operation, all I/O endpoints connected to the switch are allocated to a single upstream host. This is the classic PCIe tree topology — a CPU or server root complex connects upstream to the switch, and all downstream devices (GPUs, NVMe SSDs, NICs, FPGAs) appear as endpoints in that single host's PCIe hierarchy.

Multi-Host Operation (ExpressFabric Composable Mode): The PEX89072 extends beyond single-host operation by allowing multiple host devices to connect upstream simultaneously, with I/O endpoints dynamically allocated to different hosts via firmware commands. A single NVMe SSD can be re-assigned from one GPU compute node to another in software, without physical re-cabling — the foundation of composable infrastructure where compute, memory, and storage resources are pooled and dynamically allocated to workloads.

SR-IOV (Single Root I/O Virtualization): PCIe SR-IOV allows a single physical PCIe function to present multiple virtual functions (VFs) to a hypervisor, enabling VM-level isolation of PCIe device access. The PEX89072 passes SR-IOV traffic correctly through its switching fabric, supporting virtualized GPU and NIC sharing in cloud computing environments.

TWC (Tunneled Windows Connection): A Broadcom-proprietary low-latency host-to-host communication capability optimized for short packets. TWC creates a direct logical connection between two hosts in the same PCIe fabric, bypassing the overhead of standard PCIe transaction layers for latency-sensitive inter-host messaging — important for synchronized AI training workloads where gradient synchronization between nodes requires minimal inter-node communication latency.

CXL (Compute Express Link): CXL 1.1 and 2.0 operate over PCIe 5.0 physical layers. The PEX89072's 32 GT/s PCIe 5.0 lanes are electrically compatible with CXL-attached memory expanders and accelerators where the system architecture requires CXL connectivity through the switch fabric.

---

2.3 Hardware Root of Trust, Secure Boot, and Attestation

Enterprise and hyperscale data center deployments increasingly require hardware-based security validation of every component in the server to meet zero-trust architecture requirements. The PEX89072 implements a comprehensive security subsystem:

Hardware Root of Trust (HRoT): An immutable hardware anchor — physically implemented in on-chip ROM and one-time programmable fuses — that is the first component executed when the switch powers up. The HRoT contains a cryptographic key hierarchy that is set at manufacturing time and cannot be modified in the field, providing a tamper-resistant foundation for all subsequent security operations.

Secure Boot with Chain of Trust (CoT): The HRoT validates the digital signature of the first-stage bootloader, which in turn validates the next stage, extending the Chain of Trust through every layer of firmware and software. Any component whose signature does not match the expected value causes the boot sequence to halt, preventing the execution of unauthorized or tampered firmware.

Attestation: The PEX89072 supports cryptographic attestation — the ability to prove to a remote verifier (a data center management platform, a cloud security service, or a provisioning system) that the switch is running the expected, unmodified firmware. This is analogous to TPM (Trusted Platform Module) attestation used for server BIOS verification, but implemented at the PCIe switch level — allowing data center operators to verify the integrity of every switch in a fabric from a central management console.

---

3.0 System Topology and Application Design

3.1 AI/ML GPU Accelerator Fabric and Training Cluster

The most demanding application for the SS22-0B00-03 is as the PCIe fabric switch in an AI/ML training cluster. Modern large-language model (LLM) training requires hundreds to thousands of GPU accelerators (NVIDIA H100, AMD Instinct MI300X, Intel Gaudi 3) to communicate with high-bandwidth, low-latency interconnects. Within a single server chassis or rack unit, the PCIe switch fabric is the interconnect between:

• Upstream ports: CPU host root complex connections (x16 PCIe 5.0 links at 128 GB/s per port)
• Downstream ports: GPU accelerators (each requiring x16 PCIe 5.0 at full 128 GB/s bidirectional)
• Storage ports: NVMe SSDs for model checkpoint storage, dataset access, and intermediate tensor storage
• Network ports: PCIe-attached RDMA NICs (NVIDIA ConnectX-7, Broadcom BCM57504) for inter-node communication

With 72 lanes in a non-blocking architecture, the PEX89072 can simultaneously switch full-bandwidth traffic between multiple x16 GPU connections and storage/network endpoints — enabling the all-to-all communication pattern required by distributed training gradient aggregation without creating bottlenecks at the switch.

A representative topology for a GPU compute node: two x16 upstream ports to dual CPU root complexes + four x16 downstream ports to four GPU accelerators + two x8 ports to NVMe SSDs + two x8 ports to RDMA NICs. Total: 2×16 + 4×16 + 2×8 + 2×8 = 32+64+16+16 = 128 lanes required — exceeding the 72-lane PEX89072, so this specific topology requires the larger PEX89144 (144 lanes) or multiple PEX89072s in a cascaded fabric. The PEX89072 is ideal for rack-scale fabric switches connecting multiple server compute nodes to a shared storage tier.

---

3.2 NVMe All-Flash Array and NVMe-oF Storage Fabric

The second major application for the SS22-0B00-03 is as the fabric controller in a NVMe All-Flash Array (AFA) or NVMe-over-Fabrics (NVMe-oF) storage system:

NVMe AFA direct-attached topology: A storage controller host (PCIe root complex) connects upstream to the PEX89072. Up to 36 ports × 4 lanes (or 18 ports × 8 lanes) of NVMe SSDs connect downstream. At PCIe 5.0 x4 (256 Gb/s per NVMe drive), 36 x4 ports deliver aggregate storage bandwidth of 36 × 32 GB/s = 1,152 GB/s — exceeding even the fastest all-flash storage architectures deployed today. This allows a single PEX89072-based storage enclosure to saturate multiple 200 Gbps or 400 Gbps network connections simultaneously.

NVMe-oF fabric gateway topology: The PEX89072 connects upstream to NVMe-oF host controllers (PCIe-attached RDMA NICs running NVMe-oF target stack) and downstream to NVMe SSDs. Remote compute hosts access the SSDs over RDMA (RoCEv2 or Infiniband) through the NVMe-oF gateway. The PCIe 5.0 fabric switch eliminates the internal bandwidth bottleneck that limited earlier PCIe Gen 3/4 NVMe-oF gateway designs — the switch can forward data between any SSD and any NIC at full PCIe 5.0 rate simultaneously.

---

3.3 Tunneled Windows Connection (TWC) and Low-Latency Host Communication

TWC (Tunneled Windows Connection) is a Broadcom-proprietary feature of the PEX89000 series that provides in-band, low-latency, message-passing between hosts connected to the same PCIe fabric switch. Unlike standard PCIe transactions (which are optimized for memory-mapped I/O and DMA bulk transfers), TWC is specifically optimized for short, latency-sensitive messages — the type used for distributed computing synchronization barriers, gradient reduction notifications, and health check signals in AI training clusters.

TWC operates at near-PCIe-physical-layer latency — dramatically lower than the Ethernet-over-RoCE paths used for inter-node communication in most AI clusters. For workloads where intra-rack communication latency dominates the training step time, TWC provides a meaningful performance advantage over network-based alternatives.

In the PEX89072, TWC is implemented in hardware and managed through Broadcom's SDK — no application code changes are required for the fabric to forward TWC messages; the application simply directs messages to specific TWC window addresses that Broadcom's driver maps to remote host windows.

4.0 SS22-0B00-03 vs. Competing PCIe 5.0 Switch ICs

4.1 SS22-0B00-03 vs. PEX89088, PEX89048, PEX89032, and PEX89144

Feature	SS22-0B00-03 (PEX89072)	SS23-0B00-03 (PEX89088)	SS28-0A00-03 (PEX89048)	SS27-0A00-03 (PEX89032)	PEX89144
Total Lanes	72	88	48	32	144
Total Ports	36	44	48	32	72
Max Port Width	x16	x16	x4	x4	x16
PCIe Generation	Gen 5.0 (32 GT/s)	Gen 5.0	Gen 5.0	Gen 5.0	Gen 5.0
Aggregate BW	576 GB/s	704 GB/s	384 GB/s	256 GB/s	1,152 GB/s
ExpressFabric	Yes	Yes	Yes	Yes	Yes
Multi-Host	Yes	Yes	Yes	Yes	Yes
HRoT / Attestation	Yes	Yes	Yes	Yes	Yes
Primary Use Case	Mid-scale AI, NVMe AFA	Large AI cluster	Dense NVMe storage	Edge/entry server	Hyperscale AI fabric
Package	FCBGA-1517	FCBGA-larger	FCBGA	FCBGA	FCBGA-2000+

Selection guidance within the PEX89000 family:

• Need 72-lane balanced fabric for mid-scale AI and NVMe AFA → SS22-0B00-03 (PEX89072) — this device; the sweet spot for 4–8 GPU server designs and high-density NVMe enclosures; most widely deployed PEX89000 variant in production infrastructure
• Need highest bandwidth in a configurable wide-port design → SS23-0B00-03 (PEX89088) — 88 lanes, more ports, maximum bandwidth for the largest single-chip switch configurations
• Need maximum port count for dense x4 NVMe SSD connections → SS28-0A00-03 (PEX89048) — 48 ports optimized for x4 NVMe workloads; better for storage-density-optimized AFA enclosures
• Need lowest cost and smallest footprint for entry servers or edge computing → SS27-0A00-03 (PEX89032) — 32 lanes, smallest package, lowest power; ideal for basic PCIe fan-out in 1U servers
• Need the absolute maximum bandwidth for the largest AI fabric → PEX89144 — 144 lanes, 1,152 GB/s aggregate; used in the highest-density hyperscale GPU cluster switch designs

---

4.2 Ordering Code Suffixes: -02 vs. -03, 0A vs. 0B Explained

Broadcom's PEX89000 ordering code structure encodes important configuration and packaging information:

-02 vs. -03 suffix: The last two digits of the Broadcom SS-series ordering code indicate the packaging format and/or speed bin. For the SS22-0B00-xx series: -02 corresponds to PEX89072B0-2-DB and -03 corresponds to PEX89072B0-3-DB. Both are bulk tray format (DB = Die in Bulk). The numeric difference reflects different validated operating configurations or speed bins of the same die. For most design-in applications, Broadcom's reference design documentation specifies which suffix is appropriate — contact Broadcom or your authorized distributor to confirm the exact functional difference between -02 and -03 for your specific application.

0A vs. 0B in the package code: In the SS ordering code structure, the middle digits (0A or 0B) indicate the package revision: 0A was used for earlier PEX89000 package variants (typically the smaller 32-port and 48-port devices: SS27-0A00-xx, SS28-0A00-xx, SS29-0A00-xx), while 0B designates the production FCBGA-1517 package used by the larger 72-lane (SS22-0B00-xx) and 88-lane (SS23-0B00-xx) variants. Always verify the package designation matches your PCB footprint — 0A and 0B devices are not footprint-compatible.

---

4.3 Pricing, Availability & Authenticity

• Authorized distributors: Arrow Electronics, Avnet, Mouser, Fusion Worldwide — carry or can source SS22-0B00-03
• Typical unit pricing: ~$300–$450 USD per unit (varies significantly with market conditions; PCIe 5.0 switch ICs are premium-priced components)
• Lead time: 1–4 weeks from stocking distributors; longer for direct factory allocation during constrained supply periods
• Minimum order quantity: Typically 1 unit (cut from tray) from distributors; tray quantity is typically 70–100 units per tray
• RoHS / Halogen-Free: Yes — Broadcom's standard for production ICs
• Country of Origin: Taiwan (manufactured by TSMC on advanced process node)

For verified authentic Broadcom SS22-0B00-03 inventory with competitive pricing and volume procurement support, visit aichiplink.com — SS22-0B00-03 listing.

---

4.4 PCB Hardware Design and SDK Integration

4.5 FCBGA-1517 Package, BGA Escape Routing, and Power Integrity

The FCBGA-1517 package is a large, high-density ball grid array presenting significant PCB design challenges. Key hardware design considerations:

PCB layer count: A minimum 12-layer PCB is required for a production PEX89072 design — typically 14–16 layers are used to accommodate the PCIe 5.0 differential pair routing, power delivery network (PDN), and signal return paths simultaneously. The PCIe 5.0 lanes require tightly controlled differential impedance (85 Ω ± 10%) on dedicated signal layers with adjacent reference planes.

BGA escape routing: The FCBGA-1517 requires micro-via (µ-via) escape routing from the BGA pads to routing layers below the package. Micro-via stacks (stacked or staggered µ-vias) with 0.1 mm drill diameter are standard. PCB fabrication requires any-layer HDI (High Density Interconnect) capability — standard through-hole via PCBs cannot escape-route a 1,517-ball BGA with PCIe 5.0 signal integrity requirements.

PCIe 5.0 signal integrity: At 32 GT/s, PCIe 5.0 differential pairs must meet extremely tight insertion loss and return loss budgets. Key PCB design rules: maintain differential impedance to ±10% (85 Ω), limit trace lengths to minimize insertion loss (target < −6 dB at 16 GHz Nyquist frequency for PCIe 5.0), match differential pair intra-pair skew to < 5 mils, and implement reference plane stitching vias adjacent to every layer change in the differential pair routing.

Power delivery network (PDN): The PEX89072 has multiple power domains (core, I/O, SerDes PLL, etc.) requiring separate, low-noise voltage regulators with appropriate bulk and high-frequency decoupling. A typical implementation uses a multi-phase DCDC converter (VRM) for the core power domain, with a dense array of 100 nF + 10 nF ceramic capacitors distributed around the BGA footprint for high-frequency decoupling. The PDN must achieve sub-10 mΩ impedance at the frequencies of interest (100 MHz–3 GHz) to prevent supply noise from corrupting the SerDes PLLs.

Thermal management: With a TDP that can reach 30–60W under maximum traffic load, the PEX89072 requires an active thermal solution — either a dedicated heatsink with a server-type high-static-pressure fan, or integration into the chassis airflow design with sufficient airflow volume over the BGA package. TIM (Thermal Interface Material) between the package and heatsink must achieve < 0.5°C/W thermal resistance to keep junction temperatures below the 85°C maximum.

---

4.6 PEX89000 SDK, Management API, and Linux Driver

Broadcom provides a comprehensive PEX89000 Software Development Kit (SDK) for system integration:

SDK components:

• PlxApi (Linux API library): C-language API for PCIe fabric management, port configuration, topology discovery, event notification, and TWC window management
• PCIe fabric discovery daemon: Automatically enumerates the PCIe topology connected to the switch and presents a logical view to management software
• Firmware package: Switch firmware image, update utilities, and secure boot signing tools
• Reference host driver: Linux kernel driver for the switch management interface

Management interfaces:

• In-band management: Via the PCIe link itself — the upstream host root complex can access the switch's management registers through the normal PCIe configuration space mechanism
• Out-of-band (OOB) management: Via a dedicated SMBus/I²C interface connected to the server BMC (Baseboard Management Controller) — allows switch configuration, monitoring, and firmware update independent of the PCIe data path status. This is the recommended management path for production data center deployments where in-band management availability depends on the host OS being operational.

Topology configuration: The PEX89000 SDK allows complete programmatic configuration of: port partition assignments (which upstream host owns which downstream endpoints), multi-host fabric topology, TWC window mapping, SR-IOV virtual function allocation, and security policy (attestation certificate management, secure firmware update authorization).

Linux kernel driver support for the PEX89000 series is available from Broadcom's support portal and is also included in mainline Linux kernel versions for basic PCIe switch transparency functionality. Advanced features (ExpressFabric multi-host, TWC) require Broadcom's proprietary SDK and kernel module.

---

5.0 How to Source Authentic SS22-0B00-03 Units

High-value PCIe switch ICs like the SS22-0B00-03 are high-priority targets for counterfeiting and remarking in the secondary market. At a unit price in the $300–$450 range, the economic incentive for fraud is significant. A counterfeit or remarked SS22-0B00-03 may operate as a basic PCIe switch at low traffic levels but fail under the full 576 GB/s bandwidth load of a production AI training cluster — or may entirely lack the ExpressFabric multi-host and TWC capabilities that justify the component's selection over lower-cost alternatives.

Sourcing best practices:

• Purchase from Broadcom authorized distributors only: Arrow Electronics, Avnet, and Mouser are Broadcom-authorized distribution partners. These channels provide factory-traceable inventory with original Broadcom packaging, lot codes, and date codes — the only way to guarantee genuine silicon.
• Verify package marking: Authentic SS22-0B00-03 units carry Broadcom's laser-marked package identifier, lot code, and date code on the FCBGA package body. At this component's price point, invest in visual inspection under magnification — genuine Broadcom BGA packages use consistent, high-contrast laser marking.
• Inspect the tray and label: Broadcom's production tray packaging includes a label with the full ordering code (SS22-0B00-03 or PEX89072B0-3-DB), quantity, lot code, country of origin (Taiwan), and Broadcom's own part number barcode. Labels with inconsistent fonts, missing fields, or non-matching barcodes are red flags.
• Functional validation: At board bring-up, validate the PEX89072 using Broadcom's SDK diagnostics: (1) Confirm the PCIe device ID and revision ID register values match the expected PEX89072 values; (2) Execute the built-in BIST (Built-In Self Test) to verify SerDes lane integrity; (3) Run a bandwidth test using Broadcom's test utilities across all active PCIe 5.0 lanes to verify full 32 GT/s link speed and error-free operation.
• Request CoC for production volumes: For volume procurement, require a Certificate of Conformance with Broadcom lot traceability and factory test data before accepting delivery.

For verified authentic Broadcom SS22-0B00-03 inventory with full traceability, competitive pricing, and expert procurement support, visit aichiplink.com.

6.0 Conclusion

The SS22-0B00-03 (PEX89072B0-3-DB) represents the state of the art in PCIe switching for data center applications as of its production release — a 72-lane, 36-port, PCIe Gen 5.0 managed switch delivering up to 576 GB/s of aggregate raw bandwidth, multi-host ExpressFabric composable connectivity, hardware-enforced security through Root of Trust and attestation, ultra-low-latency TWC inter-host messaging, and SR-IOV support for virtualized environments.

For engineers designing AI/ML training cluster infrastructure, high-density NVMe All-Flash Array enclosures, NVMe-oF storage gateways, or composable rack-scale compute platforms, the SS22-0B00-03 is the correct PCIe fabric switching element where a 72-lane device is the right scale. Its position in Broadcom's PEX89000 family — between the 32-lane and 48-lane smaller siblings and the 88-lane and 144-lane larger variants — makes it the most widely deployed configuration for mid-scale AI infrastructure where the design requires more than a simple x16 fan-out but does not yet demand the full 1,152 GB/s of the PEX89144.

The hardware design complexity (FCBGA-1517, HDI PCB, multi-domain PDN, 12+ layer stackup) and software integration requirements (Broadcom SDK, firmware provisioning, secure boot management) make the SS22-0B00-03 a component reserved for experienced PCIe system architects working on production-scale data center hardware. For those engineers, it delivers a performance capability that no other single-chip commercial PCIe switch IC can match at the 72-lane scale.

Ready to source? Explore verified authentic Broadcom SS22-0B00-03 inventory with competitive pricing and full traceability at aichiplink.com — SS22-0B00-03.

Search SS22-0B00-03 Stock Now

 

 

 

 

 

Written by Jack Elliott from AIChipLink.

 

AIChipLink, one of the fastest-growing global independent electronic   components distributors in the world, offers millions of products from thousands of manufacturers, and many of our in-stock parts is available to ship same day.

 

We mainly source and distribute integrated circuit (IC) products of brands such as Broadcom, Microchip, Texas Instruments, Infineon, NXP, Analog Devices, Qualcomm, Intel, etc., which are widely used in communication & network, telecom, industrial control, new energy and automotive electronics. 

 

Empowered by AI, Linked to the Future. Get started on AIChipLink.com and submit your RFQ online today!