What Is an OAM Module? Open Accelerator Module Standard for AI Hardware

Introduction

The AI hardware industry has historically been dominated by proprietary accelerator interfaces. NVIDIA's SXM socket, for example, is an NVIDIA-defined standard that locks GPU modules, baseboards, and server chassis into NVIDIA's own ecosystem. For hyperscale cloud providers and ODMs building their own AI infrastructure, this creates supply chain constraints, limits design flexibility, and makes it difficult to mix accelerators from different vendors in a common server platform.

The Open Accelerator Module (OAM) standard was created to address exactly this problem. Developed under the Open Compute Project (OCP), OAM defines a vendor-neutral mechanical and electrical interface for AI accelerator modules, allowing different GPU and AI chip vendors to build accelerators that plug into a common baseboard design. For PCB engineers and server hardware designers, OAM introduces a standardized but highly demanding set of board-level requirements that are worth understanding in detail.

1. Table of Contents
2. Introduction
3. What Is an OAM Module?
4. OCP and the Open Accelerator Infrastructure Project
5. Why OAM Exists: The Problem with Proprietary AI Accelerator Interfaces
6. OAM Architecture
7. - Module Types: OAM and UBB
8. - Electrical Interface
9. - Mechanical Form Factor
10. - Power Delivery
11. - Thermal Interface
12. OAM vs SXM: Open Standard vs NVIDIA Proprietary
13. Who Uses the OAM Standard?
14. PCB Design Requirements for OAM-Compatible Systems
15. - Universal Base Board PCB Design
16. - Signal Integrity on UBB
17. - Power Delivery on UBB
18. - Thermal Management on UBB
19. - OAM Module PCB Design
20. OAM Standard Generations
21. Manufacturing OAM Modules and UBBs
22. FAQ

---

What Is an OAM Module?

An OAM (Open Accelerator Module) is a standardized form-factor AI accelerator module defined by the Open Compute Project's Open Accelerator Infrastructure (OAI) workgroup. An OAM module contains one or more AI accelerator dies (GPU, NPU, or custom ASIC), high-bandwidth memory (typically HBM), power management components, and a standardized edge connector that plugs into a Universal Base Board (UBB).

The OAM standard defines:

• Physical dimensions and mechanical envelope of the module
• Edge connector pinout and signal assignment
• Electrical interface specifications (PCIe, high-speed serial links, power rails)
• Thermal interface requirements (contact area, flatness tolerance)
• Power delivery specifications (voltage levels, current limits, sequencing)
• Management interface (I²C, SMBus, MCTP)

The complementary component is the Universal Base Board (UBB)—the PCB into which OAM modules plug. The UBB provides power, PCIe connectivity to the host CPU, high-speed inter-accelerator links, and thermal management infrastructure. Because the UBB interface is standardized, a UBB designed to the OAM specification can in principle accept OAM modules from any compliant vendor.

---

OCP and the Open Accelerator Infrastructure Project

The Open Compute Project (OCP) is a collaborative engineering community founded by Facebook (now Meta) in 2011, focused on developing open-source hardware standards for data center infrastructure. OCP contributions cover server chassis, rack architecture, network switches, storage, and increasingly, AI accelerator hardware.

The Open Accelerator Infrastructure (OAI) project within OCP was established to define open standards for AI accelerator modules and their host boards. The OAM specification emerged from OAI as the primary module standard, with contributions from AMD, Intel, Google, Qualcomm, Graphcore, and several ODMs and system integrators. NVIDIA has not adopted the OAM standard for its primary data center GPU products, preferring its proprietary SXM interface.

OCP specifications are publicly available and royalty-free, enabling any hardware vendor to design OAM-compliant modules or UBBs without licensing fees.

---

Why OAM Exists: The Problem with Proprietary AI Accelerator Interfaces

Before OAM, each major AI accelerator vendor defined its own module form factor and baseboard interface:

• NVIDIA SXM (SXM4 for A100, SXM5 for H100, SXM6 for B200): NVIDIA-proprietary, requires NVIDIA-licensed baseboard design
• AMD OAM-compatible modules (MI250X, MI300X): AMD adopted OAM for its server GPU line
• Google TPU modules: Custom form factors specific to Google's TPU pods
• Graphcore IPU: Custom form factors

The proprietary approach creates several problems for cloud providers and ODMs:

• Single-vendor lock-in: A server designed around NVIDIA SXM cannot accept AMD or Intel accelerator modules; the entire baseboard must be redesigned when switching accelerator vendors
• Longer time-to-market: Each new GPU generation may introduce a new proprietary socket, requiring a full baseboard redesign even for the same vendor's products (SXM4 → SXM5 → SXM6 are all incompatible)
• Supply chain concentration: With a single accelerator source per platform, supply disruptions have outsized impact
• Innovation constraints: ODMs cannot differentiate their server designs at the module interface level

OAM addresses these problems by separating the module design (accelerator vendor's responsibility) from the baseboard design (ODM's or cloud provider's responsibility) through a published, vendor-neutral interface specification.

---

OAM Architecture

Module Types: OAM and UBB

The OAM ecosystem consists of two complementary components:

Component	Description	Designed By
OAM Module	The accelerator module containing AI chip(s), HBM, and power management; plugs into UBB via edge connector	Accelerator vendor (AMD, Intel, etc.)
Universal Base Board (UBB)	The host PCB that accepts up to 8 OAM modules; provides power, PCIe, inter-module links, and management	ODM, cloud provider, or system integrator
OAM Carrier / Server Chassis	Rack-mount enclosure housing the UBB, power supplies, and cooling infrastructure	ODM or cloud provider

The OAM specification also defines two module size classes:

• OAM (standard): Approximately 74 mm × 74 mm module footprint; designed for single-die or moderate-density accelerators
• OAM-XL (extended): Larger footprint for higher-power, multi-die accelerators requiring more PCB area and higher thermal dissipation

Electrical Interface

The OAM edge connector carries the following signal categories:

Signal Category	Interface	Notes
Host interface (CPU to accelerator)	PCIe Gen4 or Gen5 ×16	Primary data path from host CPU; Gen5 in current generation
Inter-module high-speed links	Vendor-defined high-speed serial (e.g., AMD Infinity Fabric)	Enables accelerator-to-accelerator communication through UBB; equivalent function to NVLink in NVIDIA systems
Management interface	I2C, SMBus, MCTP over PCIe	Out-of-band management: power sequencing, thermal monitoring, firmware update
Power rails	12 V (legacy), 48 V (preferred in OAM Gen2+)	High-current power delivered from UBB power planes through edge connector pins
Auxiliary signals	GPIO, UART, reset, presence detect	System management and hot-plug support

The edge connector in the OAM specification uses a high-density, high-current connector capable of handling the combined signal and power requirements of a 700–1,000 W accelerator module. Connector contact pitch is in the 0.5–0.8 mm range depending on signal type, with power pins sized for the required current capacity.

Mechanical Form Factor

OAM module mechanical dimensions are tightly specified to ensure compatibility across vendors and UBB designs:

• Standard OAM footprint: approximately 74 mm (width) × 74 mm (depth)
• Module height (above UBB surface): specified maximum to ensure heatsink or cold plate clearance within the chassis
• Edge connector location: defined precisely relative to module outline; connector mating is perpendicular to the UBB surface (vertical insertion)
• Keepout zones: defined around the connector area and module perimeter to ensure mechanical clearance between adjacent modules and chassis structures
• Flatness tolerance on thermal interface surface: < 0.1 mm across the module top surface to ensure adequate thermal contact with heatsink or cold plate

A standard UBB accepts up to 8 OAM modules arranged in two rows of four, with defined module-to-module pitch to maintain airflow channels or cold plate alignment.

Power Delivery

OAM modules in current AI accelerator applications consume 400–750 W per module. The OAM specification supports two primary power architectures:

• 12 V bus (OAM Gen1): 12 V delivered from the UBB to the module; on-module VRMs convert to the required core and I/O voltages. At 700 W, this requires approximately 58 A at 12 V through the edge connector power pins—a significant current density challenge for the connector and UBB power planes
• 48 V bus (OAM Gen2+): 48 V delivered from the UBB to the module; on-module buck converters step down to core voltages. At 700 W, this requires approximately 14.6 A at 48 V—a 4× reduction in current for the same power, significantly relaxing connector and plane sizing requirements. The 48 V architecture is preferred for high-power OAM modules

Thermal Interface

OAM modules interface with the chassis thermal management system through a defined contact area on the top surface of the module. The specification defines:

• Contact area dimensions and location relative to module outline
• Flatness tolerance (< 0.1 mm) to ensure TIM (thermal interface material) contact across the full area
• Maximum operating temperature at the module thermal interface surface
• Minimum thermal conductivity requirement for TIM between module and heatsink/cold plate

OAM-compliant chassis may use air cooling (heatsinks with forced airflow) or direct liquid cooling (cold plates). At 700 W+ per module and 8 modules per UBB (5,600 W+ total), direct liquid cooling is strongly preferred for sustained operation.

---

OAM vs SXM: Open Standard vs NVIDIA Proprietary

Parameter	OAM (Open Standard)	NVIDIA SXM (Proprietary)
Governing body	Open Compute Project (OCP)	NVIDIA
Specification access	Public, royalty-free	Licensed; NDA required for full spec
Compatible accelerators	AMD MI series, Intel Gaudi, others	NVIDIA GPUs only (A100, H100, B200)
UBB / baseboard design	Any ODM or cloud provider can design a compliant UBB	NVIDIA-designed or NVIDIA-licensed HGX baseboard
Inter-module interconnect	Vendor-defined (AMD Infinity Fabric, etc.); routed on UBB	NVLink 4.0 / 5.0; routed on NVIDIA baseboard
Host interface	PCIe Gen4 / Gen5 (standardized)	PCIe Gen5 (SXM5) / PCIe Gen6 (SXM6)
Power bus	12 V (Gen1) or 48 V (Gen2+)	12 V (SXM4/5) / 48 V (SXM6 / GB200)
Module generations	OAM Gen1, Gen2; evolving	SXM4 (A100), SXM5 (H100/H200), SXM6 (B200)
Cross-generation compatibility	OAM Gen1 and Gen2 maintain connector compatibility (with caveats)	SXM4, SXM5, SXM6 are mutually incompatible
Primary users	AMD MI300X deployments, cloud providers building open AI servers	DGX, HGX, and OEM NVIDIA AI server programs

---

Who Uses the OAM Standard?

OAM adoption is concentrated among cloud providers seeking supply chain flexibility and accelerator vendors who are not NVIDIA:

• AMD: The MI250X, MI300X, and MI350 accelerators are all shipped in OAM-compatible form factors. AMD's OAM modules slot into OAM-compliant UBBs designed by ODMs or cloud providers, making AMD the primary GPU vendor implementing OAM at scale
• Intel: Intel Gaudi 2 and Gaudi 3 AI accelerators use an OAM-derived module form factor, enabling deployment in OAM-compatible infrastructure
• Meta, Microsoft, Google: These hyperscale cloud providers have contributed to OAM specification development and deploy OAM-based infrastructure for workloads where they are not locked to NVIDIA SXM systems
• ODMs (Wiwynn, Quanta, Inventec, Foxconn): Major ODMs design OAM-compliant UBBs and server chassis that can be populated with OAM modules from multiple accelerator vendors

NVIDIA does not currently ship products in OAM form factor. NVIDIA's equivalent of the OAM ecosystem is its HGX platform, which uses NVIDIA-proprietary SXM sockets and NVIDIA-designed or NVIDIA-licensed baseboards.

---

PCB Design Requirements for OAM-Compatible Systems

Universal Base Board PCB Design

The UBB is the most complex PCB in an OAM-based AI server. A UBB designed for 8 OAM modules must simultaneously accommodate:

• 8 high-density OAM edge connectors with power and signal contacts
• PCIe Gen5 routing from host CPU(s) to each OAM slot
• High-speed inter-module serial links (AMD Infinity Fabric or equivalent) routed between module slots
• 48 V or 12 V power distribution planes with on-board VRMs or power conversion stages
• Management network (I²C / SMBus) connecting to all 8 module slots and to the server BMC
• Thermal monitoring (temperature sensors, fan control or liquid cooling management)

Typical UBB layer counts range from 16 to 28 layers, depending on the inter-module interconnect bandwidth and PCIe generation. UBBs designed for AMD MI300X with Infinity Fabric inter-module links and PCIe Gen5 host interfaces commonly use 20–24 layers.

Signal Integrity on UBB

The UBB must route high-speed signals across relatively long board distances (200–400 mm from one module slot to another in an 8-slot configuration). Signal integrity requirements:

• PCIe Gen5 (×16 per slot): 32 GT/s per lane; channel insertion loss < 28 dB at 16 GHz; backdrilling required on all through-hole vias in PCIe signal paths
• AMD Infinity Fabric inter-module links: Up to 128 GB/s bidirectional per link; differential pair impedance 100 Ω ± 5%; low-loss laminate required on these routing layers
• OAM edge connector launch: The transition from PCB trace to edge connector introduces an impedance discontinuity that must be minimized through pad geometry and ground plane design; connector vendor's IBIS models should be used in channel simulation
• Length matching: All lanes within a PCIe or Infinity Fabric bundle matched to < 5 ps intra-pair skew; bundle-to-bundle matching per specification requirements

Power Delivery on UBB

At 700 W per module × 8 modules, a fully loaded UBB must deliver up to 5,600 W of accelerator power. Power delivery architecture on the UBB:

• 48 V input bus: Preferred for high-power OAM Gen2+ systems; 48 V enters the UBB from the server PSU and is distributed across dedicated 48 V power planes to each module connector; at 5,600 W total, the 48 V bus carries approximately 117 A—requiring 3 oz or heavier copper on the 48 V planes and bus bar or heavy copper trace distribution within the board
• 12 V input bus (legacy): At 5,600 W, 12 V bus current exceeds 466 A; this is only practical with multiple parallel power paths and is generally avoided in new high-power designs
• On-board VRMs (if present): Some UBB designs include on-board VRMs to pre-regulate voltage before the module connector, reducing the current through the connector power contacts; VRM selection and placement follows the same rules as GPU baseboard PDN design
• Power plane copper thickness: 3–4 oz copper (105–140 μm) on primary power planes; heavier copper increases plane resistance and requires modified etching parameters during fabrication

Thermal Management on UBB

The UBB itself operates in a high-temperature environment (8 OAM modules dissipating 700 W each produce significant radiant heat). Board-level thermal requirements:

• Thermal vias under connector areas transfer heat from power-carrying connector pads to internal copper planes, reducing connector pad temperature
• Board material T_g ≥ 170°C; sustained operation in a 5,600 W module environment elevates board temperature significantly, and lower-T_g materials risk delamination over time
• Copper plane area under module connectors must be maximized to act as a heat spreader, distributing thermal load from connector power pins across a wider board area before it reaches the plane's edge or via structures

OAM Module PCB Design

The OAM module itself is a highly complex PCB assembly within the standardized mechanical envelope. Module PCB design requirements depend on the specific accelerator die(s) and HBM configuration, but common characteristics include:

• Layer count: 12–20 layers within the module; the module PCB is smaller than a UBB but must route power and signals from the edge connector to large BGA packages in a confined area
• Substrate technology: For modules with very fine-pitch accelerator packages (e.g., AMD MI300X using CoWoS or equivalent packaging), module PCBs may approach IC substrate territory with trace widths and via sizes at the lower end of standard PCB capability
• Edge connector footprint: The OAM edge connector on the module bottom is a precision feature—contact pitch, gold plating thickness, and coplanarity across all contacts must be tightly controlled to ensure reliable mating with the UBB connector and low contact resistance for high-current power pins
• Thermal interface surface: The top surface of the module PCB (or a metal spreader mounted to it) must meet the flatness specification (< 0.1 mm) of the OAM standard; post-assembly warpage measurement and management is part of the OAM module manufacturing process

---

OAM Standard Generations

Parameter	OAM Gen1	OAM Gen2
Host interface	PCIe Gen4 ×16	PCIe Gen5 ×16
Primary power bus	12 V	48 V (preferred)
Module max TDP	~400 W	~700 W+ (roadmap to 1,000 W)
Inter-module links	Vendor-defined; lower bandwidth	Vendor-defined; higher bandwidth (e.g., Infinity Fabric Gen4)
Management	I2C / SMBus	MCTP over PCIe + I2C / SMBus
Connector backward compatibility	Gen1 connector baseline	Mechanically compatible with Gen1 slot (with power caveats)
Representative products	AMD MI250X	AMD MI300X, Intel Gaudi 3

---

Manufacturing OAM Modules and UBBs

Both OAM modules and UBBs require PCB manufacturing capabilities well above commodity standards:

OAM Module Manufacturing:

• Fine-pitch BGA assembly for AI accelerator packages (CoWoS or flip-chip); solder joint inspection by X-ray is mandatory
• Edge connector gold plating to OAM specification (typically ENIG or hard gold on contact fingers); thickness uniformity and adhesion critical for reliable mating over thousands of insertion cycles
• Post-assembly flatness measurement and warpage correction; modules that exceed the 0.1 mm flatness tolerance must be reworked or scrapped
• Burn-in and functional testing at module level before integration; AI accelerator packages are high-value components and yield management at the module assembly stage is essential

UBB Manufacturing:

• Large-format PCB fabrication (UBBs for 8 OAM modules can exceed 500 mm × 600 mm)
• High-layer-count fabrication with hybrid stackup (low-loss laminates on high-speed signal layers; standard materials on power and management layers)
• Heavy copper processing (3–4 oz) on 48 V power planes; heavy copper requires modified etching parameters and careful press cycle management to maintain dielectric thickness uniformity
• Controlled-depth backdrilling on PCIe Gen5 and high-speed inter-module signal vias
• OAM connector assembly: high-density, high-pin-count connectors require precise placement and controlled-force press-fit or soldering processes; coplanarity of all connector housings across 8 slots must be maintained to ensure simultaneous module engagement

---

FAQ

Does NVIDIA support the OAM standard?
No. NVIDIA uses its proprietary SXM form factor (SXM4, SXM5, SXM6) for its data center GPU modules. NVIDIA has not adopted the OAM standard for its H100 or B200 GPU lines. NVIDIA's HGX platform is the proprietary equivalent of the OAM ecosystem—a standardized baseboard and module system, but controlled by NVIDIA rather than an open standards body.

What is the difference between OAM and HGX?
OAM is an open standard governed by the Open Compute Project, usable by any accelerator vendor. HGX is NVIDIA's proprietary equivalent—a baseboard and module specification controlled by NVIDIA that supports only NVIDIA GPUs. Both achieve a similar goal (standardizing the accelerator-to-baseboard interface within a vendor's ecosystem), but OAM does so across multiple vendors while HGX is NVIDIA-exclusive.

Can an AMD MI300X OAM module be used in an NVIDIA HGX baseboard?
No. OAM and HGX/SXM are physically and electrically incompatible. An AMD MI300X OAM module requires an OAM-compliant UBB; it cannot be installed in an NVIDIA SXM5 or SXM6 socket.

How many OAM modules fit in a standard server?
A standard OAM UBB accommodates 8 OAM modules. A 2U server chassis can typically house one UBB (8 modules), making 8 accelerators per server the common configuration for OAM-based AI servers.

What is the maximum power per OAM module?
OAM Gen1 supports up to approximately 400 W per module. OAM Gen2 supports up to 700 W per module today, with the specification roadmap extending toward 1,000 W per module for future accelerator generations. The 48 V power bus in OAM Gen2 is essential for managing the high currents at these power levels.

Is OAM used for inference or training workloads?
OAM modules are used for both training and inference. AMD's MI300X, which ships in OAM form factor, is deployed for both large-scale training (competing with H100) and high-throughput inference. The OAM standard itself does not specify the workload—it defines only the physical and electrical interface between module and baseboard.

---

Need to Manufacture AI Server PCBs?

OAM modules and Universal Base Boards demand precision fabrication: large-format PCBs, heavy copper power planes, low-loss laminates, fine-pitch BGA assembly, and rigorous post-assembly inspection. NextPCB's advanced PCB manufacturing capabilities support the full OAM hardware stack—from prototype module PCBs to production UBBs.

 

---

Related Articles:

• What Is an AI Server? Architecture, Components & PCB Requirements
• NVIDIA Blackwell Architecture Explained: B200, GB200 & PCB Design Impact
• What Is NVLink? How NVIDIA's High-Speed GPU Interconnect Shapes PCB Routing
• What Is NVSwitch? The Silicon Behind NVIDIA's GPU Cluster Scale-Out
• A100 vs H100: GPU Generational Leap & PCB Stack Differences Explained