PEX90144 - 144-Lane PCIe Gen6 Switch | Broadcom | AI
MPN: PEX90144 ✓ Active| Qty | Unit Price | Extended |
|---|---|---|
| 1 | $0 | $0.00 |
| 10 | $0 | $0.00 |
| 100 | $0 | $0.00 |
| 500 | $0 | $0.00 |
| 1,000 | $0 | $0.00 |
PEX90144 Overview
A PCIe switch is a fabric device that expands a single root-complex PCI Express link into multiple downstream endpoints, aggregating GPU, NIC, storage controller, and accelerator traffic. In the system hierarchy, it sits between the host CPU root port and the endpoint tier, and is a core building block of modern data center interconnect fabrics alongside CPU PCIe controllers and retimers.
Key differentiating features include hardware multi-host support for sharing the switch fabric across several hosts, an in-house Broadcom SerDes PHY exceeding 36dB of performance margin at Gen6 PAM-less NRZ data rates, and an embedded PCIe protocol analyzer for in-system debug without external interposer hardware. Turnkey fabric management software simplifies bring-up and topology configuration for large racks.
Architecturally, the 5nm CMOS process integrates the SerDes, packet-routing fabric, and management engines on a single die, keeping per-lane power low while sustaining Gen6 64GT/s signaling. The device scales above its siblings, sitting as the widest device in Broadcom's published PCIe Gen6 switch lineup, which also includes 104-lane and 88-lane configurations.
Typical applications include AI rack-scale connectivity between host CPUs and GPU accelerators, high-density server PCIe expansion chassis, composable and disaggregated infrastructure, and NVMe storage fabrics. It was publicly demonstrated integrated into a Dell PowerEdge XE7745-style PCIe board at SC25.
Design consideration: at this bandwidth scale, signal-integrity planning for Gen6 64GT/s channels, substrate stack-up, and thermal design for the large BGA package are critical; consult the Broadcom PEX90000 series product brief for validated topologies.
This page synthesizes distributor availability, family cross-references, and practical design guidance not found in the manufacturer datasheet.
Drop-in alternatives for PEX90144 — same package, pin-to-pin compatible. Different-package parts requiring PCB rework are excluded.
Quick Comparison Tool — Select alternative parts for side-by-side comparison:
No drop-in alternatives available for this product.
Request AlternativesPEX90144 Maximum Ratings & Electrical Characteristics
| Product Type | PCI Express 6.0 Packet Switch |
| Total Lanes | 144 |
| PCIe Generation | Gen 6.0 |
| Supported Port Widths | x2, x4, x8, x16 |
| Bidirectional Raw Bandwidth per x16 Port | 256 GB/s |
| Maximum Raw Bandwidth Through Device | 9.2 Tb/s |
| Process Technology | 5 nm CMOS |
| SerDes Performance | >36 dB (in-house SerDes) |
| Multi-Host Support | Yes (hardware) |
| Embedded Analyzer | Yes (embedded PCIe analyzer) |
| Fabric Management | Turnkey fabric management software |
| Family | PEX90000 Series (ExpressFabric) |
PEX90144 standard Pin Configuration Guide
Pin configuration for PEX90144 (standard package). This analog component features input, output, and power supply pins. Refer to the manufacturer datasheet for offset null, compensation, and enable pin configurations. Ideal for signal conditioning and amplifier circuits.
No detailed pinout data available for PEX90144.
Refer to the datasheet for full pin configuration.
Typical Applications
PEX90144 is suitable for 6 applications: AI Rack-Scale GPU Connectivity, High-Density Server PCIe Expansion, NVMe Storage Fabrics, Composable and Disaggregated Infrastructure, Test, Measurement and Protocol Validation Platforms, Next-Generation Network Appliance Backplanes.
AI Rack-Scale GPU Connectivity
The PEX90144 is purpose-built for AI rack connectivity, where a host CPU must fan out PCIe Gen6 links to many GPU accelerators. Its 144 lanes can be carved into nine x16 ports at 256GB/s bidirectional raw bandwidth each, providing the fan-out density required for 8-GPU and higher accelerator pods. The 5nm process keeps per-lane power low, and hardware multi-host support allows multiple hosts to share the fabric, which is essential for resource-pooled AI training clusters. QS Compute and Broadcom position the device as the widest Gen6 switch in the published lineup for exactly this role, with turnkey fabric management software accelerating rack bring-up versus custom firmware development.
Recommended
High-Density Server PCIe Expansion
In 2U/4U expansion chassis, the PEX90144 aggregates a narrow upstream link from the host into many downstream slots for GPUs, NICs, and storage HBAs. At PCIe Gen6, a single x16 upstream sustains 256GB/s bidirectional raw bandwidth, and the switch's x2/x4/x8/x16 port carving lets designers match downstream widths to slot types, avoiding stranded lanes. The embedded PCIe analyzer simplifies in-chassis link debugging where physical access to interposers is impractical. Broadcom demonstrated the device in a Dell PowerEdge XE7745-style PCIe board at SC25, validating its fit in dense x86 server expansion topologies with standard fabric management software handling enumeration and hot-plug behavior.
Recommended
NVMe Storage Fabrics
The PEX90144 enables large-scale NVMe disaggregation by connecting dozens of NVMe drives to host complexes through Gen6 links. Its 144 lanes can fan out as many x4 drive ports, each running at Gen6 rates so that a single switch feeds the maximum number of SSDs per host lane budget. Because storage latency budgets are tight, the switch's low-hop architecture and embedded PCIe analyzer help verify end-to-end link quality and diagnose latency anomalies during system integration. Compared with cascading multiple smaller Gen4 switches, one PEX90144 reduces hop count, board complexity, and per-port cost, making it attractive for composable storage and JBOF controller designs in hyperscale data centers.
Recommended
Composable and Disaggregated Infrastructure
Composable infrastructure requires dynamically reassigning PCIe endpoints among hosts, and the PEX90144's hardware multi-host support is a direct architectural fit: multiple host CPUs can each own segments of the 144-lane fabric, with turnkey fabric management software reconfiguring topology at runtime. Per the Broadcom PEX90000 product brief, the family's up-to-9.2Tb/s raw switching bandwidth means resource pools of GPUs, NICs, and storage can be re-composed without oversubscribing the fabric. The embedded analyzer further supports the validation cycles that dynamic recomposition demands, catching link-training regressions after each topology change. This makes the device a core building block for software-defined data center fabrics.
Recommended
Test, Measurement and Protocol Validation Platforms
Because the PEX90144 integrates an embedded PCIe protocol analyzer, it doubles as a validation platform for Gen6 endpoint bring-up: engineers can capture link-level traffic on internal lanes without external interposer hardware, which is especially valuable at Gen6 64GT/s signaling where interposers degrade channel margin. The in-house SerDes rated above 36dB provides headroom for extended channels typical of test racks. System validation labs can use the 144-lane density to test many endpoints in parallel - for example, qualifying NVMe drives, GPUs, or NICs at scale - while the fabric management software automates enumeration scenarios and error-injection test cases across the full port matrix.
Recommended
Next-Generation Network Appliance Backplanes
High-throughput network appliances - 400G/800G security gateways, load balancers, and SDN forwarders - need to aggregate many smart NICs behind limited host lanes. The PEX90144 addresses this with nine possible x16 Gen6 ports at 256GB/s bidirectional raw bandwidth each, enough to feed multiple 800G-class NICs without oversubscription. Hardware multi-host support lets redundancy controllers share the same NIC pool for high-availability designs. ServeTheHome's SC25 coverage notes the device was shown in production-style PCIe boards, indicating ecosystem readiness. Compared with the 104-lane PEX90104, the extra 40 lanes translate directly into more NIC slots per appliance, improving per-rack throughput density.
Recommended
Recommended Products Summary
Engineering reference data for PEX90144 — comparison, design guidance, and compliance information.
Selection Guide
Comparison with Alternatives
| Parameter | This Product | PEX90104 | PEX9700 |
|---|---|---|---|
| PCIe Generation | Gen 6.0 | Gen 6.0 | Legacy ExpressFabric generation |
| Multi-Host Support | Yes (hardware) | Yes (hardware) | Managed switch family support |
| Brand | Broadcom | Broadcom | Broadcom (Avago Technologies) |
Key Differentiators
- Widest lane count in the published Gen6 lineup (vs PEX90104)
- 5nm process efficiency (vs PEX9700)
- Integrated debug and management (vs PEX90104)
Design Notes
PCIe Gen6 operates at 64GT/s, so every channel on the PEX90144 demands rigorous signal-integrity engineering: control insertion loss within the PCIe 6.0 budget, use low-loss laminate, minimize via stubs via back-drilling, and simulate channels including the package model. The in-house Broadcom SerDes is rated above 36dB, but designers should not assume the margin transfers to their board; validate with the PEX90000 series reference topologies. At 144 lanes, plan the substrate stack-up (typically high layer count) early, since routing density of the BGA breakout is frequently the dominant board cost driver.
A 144-lane 5nm Gen6 switch concentrates significant power in a single large BGA. Although exact typical power is not public, the same class of device in AI switch boards is normally fitted with a bonded heatsink and forced airflow in server chassis. Perform a junction temperature analysis at maximum lane utilization and worst-case ambient per your chassis airflow model before layout freeze, and coordinate with Broadcom for the qualified operating range (not stated in the public product brief). Estimated: verify thermal design against measured power data from Broadcom under NDA rather than extrapolating from smaller family members.
Do not treat the PEX90144 as a footprint replacement for the PEX90104: the lane counts (144 vs 104) and package differ, so board respins are required. Use Broadcom turnkey fabric management software rather than writing custom switch firmware, which shortens bring-up and preserves the embedded analyzer functionality. For multi-host designs, agree on host ownership and hot-reset behavior early - hardware multi-host support exists, but host OS enumeration strategy still requires careful planning to avoid contention over shared endpoints.
Compliance Information
Compliance details for the PEX90144 were not stated in the provided public web data; the PEX90000 series product brief does not include environmental compliance tables. Confirm RoHS/REACH status via Broadcom product compliance documentation.