Surface
Stacy Lu
NextPCB Capabilities
Printed Circuit Boards
PCB Assembly
Layer Buildup
SMD-Stencils
PCB Design-Aid & Layout
Mechanics
Quality
Drills & Throughplating
Factory & Certificate
PCB Assembly Factory Show
Certificate
Support Team
Feedback:
support@nextpcb.com
An OAM (Open Accelerator Module) board is one of the most demanding PCB assembly jobs in commercial production. Inside a standardized mechanical envelope roughly 74 mm × 74 mm, the module carries one or more AI accelerator dies with hundreds of billions of transistors, multiple HBM memory stacks, power management ICs, decoupling capacitors, and a high-density edge connector that must mate reliably with a Universal Base Board (UBB) and deliver up to 750 W of power while carrying PCIe Gen5 ×16 and Infinity Fabric signals at multi-gigabit speeds.
Getting every step of the manufacturing process right—from bare board fabrication through final functional test—determines whether the module performs to specification, survives thermal cycling in a data center, and can be produced at acceptable yield. This guide walks through the complete OAM module and UBB manufacturing process in the order it occurs on the production floor, with specific attention to the steps that differ from standard PCB assembly and where failures most commonly originate.
For background on the OAM standard itself—its electrical interface, mechanical specification, and relationship to the NVIDIA SXM ecosystem—the OAM Module introduction provides the foundational context before diving into manufacturing.
The OAM standard, defined by the Open Compute Project (OCP), specifies a common mechanical and electrical interface for AI accelerator modules. An OAM module contains the accelerator die(s) and HBM memory stacks, and plugs into a Universal Base Board (UBB) via a high-density edge connector. AMD MI300X, Intel Gaudi 3, and several custom AI chips ship in OAM form factor. The UBB provides power distribution, PCIe Gen5 host connectivity, Infinity Fabric inter-module routing, and liquid cooling infrastructure for up to 8 modules simultaneously.
Manufacturing complexity matters in this context because OAM modules concentrate several of the most difficult PCB assembly challenges into a single small board:
Any one of these challenges is demanding in isolation. All of them on a single small PCB, manufactured to IPC Class 3 standards, is what makes OAM module production one of the highest-complexity PCB assembly jobs in the AI hardware supply chain.
Before a single panel enters the production floor, the manufacturing engineering team reviews the complete design package and generates the production documentation. For OAM modules, this review is more extensive than for standard PCB assemblies because of the tight tolerances and specialized process steps involved.
The design package for an OAM module should include: Gerber or ODB++ fabrication data with the complete layer stackup specification; drill files including laser drill files for microvia layers and controlled-depth backdrill files for through-hole vias on high-speed signal layers; the edge connector gold plating specification (thickness, alloy, and coverage area); the thermal interface flatness requirement and the method for verification; the accelerator package thermal reflow profile or reference to the component vendor's published profile requirements; and the functional test specification including power-on sequence, PCIe enumeration check, memory bandwidth benchmark, and burn-in duration.
The DFM (Design for Manufacturability) review specifically checks: whether via-in-pad structures are fully specified with fill material (conductive or non-conductive epoxy) and cap-plating callout; whether the edge connector footprint geometry matches the connector vendor's recommended land pattern at the required gold plating thickness; whether BGA pad sizes are consistent with the NSMD (Non-Solder-Mask-Defined) or SMD specification required by the accelerator package vendor; and whether thermal relief geometries on power connector pads are appropriate for the reflow profile.
Stackup definition for OAM module PCBs is reviewed against the signal integrity requirements of the specific AI accelerator package. For AMD MI300X OAM modules, the UBB carries PCIe Gen5 and Infinity Fabric signals requiring Megtron 6E or Tachyon 100G on signal layers, as discussed in the high-speed PCB materials guide. The module PCB itself routes power, management, and HBM-adjacent signals from the edge connector to the accelerator package, which may have less demanding laminate requirements than the UBB but still demands careful impedance control on any high-frequency management signals.
OAM module PCBs are compact (approximately 74 mm × 74 mm) but high-complexity boards. Typical characteristics: 12–20 layers, HDI build-up structure (1+N+1 minimum, 2+N+2 for fine-pitch accelerator packages), via-in-pad on accelerator and HBM package pads, controlled impedance on PCIe and management signal layers, and a precision edge connector area with hard gold plating.
Stackup lamination: OAM module PCBs with 2+N+2 HDI require three lamination press cycles. The core is pressed first from inner layer pairs and prepreg; then the first build-up layers are added on each side and laser-drilled; then the second build-up layers complete the HDI structure. Each press cycle uses the laminate and prepreg materials specified for the module design—typically Megtron 6 or equivalent on power and non-critical layers, with lower-loss material on any layer carrying high-frequency signals from the edge connector. The thermal interface surface of the module is ultimately defined by the mechanical flatness of the finished PCB assembly, not the bare board alone; however, gross PCB warpage at the bare board stage (exceeding approximately 0.3% of panel diagonal) will propagate through assembly and must be corrected at the fabrication stage.
Laser drilling: Microvias on OAM module PCBs must be drilled to < 100 μm diameter for fine-pitch accelerator package BGA escape. CO2 laser drilling parameters are qualified for the specific prepreg material used in the build-up layers. Stacked microvia alignment (for stacked via structures connecting build-up layers) must be within ± 25 μm of nominal to ensure adequate land coverage at each layer interface.
Via-in-pad processing: The accelerator package BGA pads and HBM package BGA pads require via-in-pad (VIPPO). The via fill process—epoxy injection, cure, grinding to planar, cap-plating—must be executed with tight dimensional control. The finished via-in-pad surface must be within < 10 μm of the surrounding pad level; a dimple or protrusion beyond this tolerance reduces solder paste volume or contact area and risks a weak solder joint under the accelerator package.
Edge connector area fabrication: The OAM edge connector contacts are part of the PCB itself—they are the exposed copper fingers at the board edge. These fingers are beveled (typically 30–45° chamfer) during board routing to facilitate insertion into the UBB connector receptacle. Before beveling, the finger area receives electrolytic nickel (3–6 μm) and hard gold (0.76–1.27 μm minimum, typically to the OAM standard specification) plating. Gold thickness and uniformity are verified by XRF (X-ray fluorescence) spectrometry at multiple points along the finger array; non-uniform plating causes contact resistance variation that degrades the high-current power pin reliability.
Bare board electrical testing: Flying probe testing verifies all inner layer continuity and isolation at 100% panel coverage. TDR (Time Domain Reflectometry) testing on impedance coupons verifies that controlled impedance structures are within the specified tolerance for any high-frequency signal layers.
The UBB is a substantially larger and more complex PCB than the individual OAM modules it hosts. A standard 8-slot UBB measures approximately 500–600 mm × 400–500 mm and typically uses 16–24 layers with hybrid low-loss stackup, heavy copper power planes, and OAM slot connector footprints at precise mechanical locations.
Large-format fabrication: UBBs exceed the standard 610 mm × 457 mm PCB panel size used in most high-volume fabrication. Large-format fabrication requires specialized lamination presses, longer drill travel on CNC equipment, and AOI systems capable of handling the larger panel size without registration drift. Fabricators without large-format capability cannot produce full-size UBBs; panel sizing constraints sometimes require UBBs to be fabricated as two panels that are joined in the assembly stage, though single-panel designs are preferred for reliability.
Heavy copper power planes: OAM UBBs operating at 48 V with up to 8 modules at 750 W each (total 6,000 W, approximately 125 A at 48 V) require power planes with 2–3 oz copper (70–105 μm) to maintain acceptable resistance and voltage drop. Heavy copper plating requires modified etch chemistry and longer etch time, and the increased copper weight on inner layers affects dielectric thickness after pressing (as copper area displaces resin flow during lamination). The fabricator's stackup design must account for this copper displacement to maintain specified dielectric thickness on adjacent signal layers.
Backdrilling: PCIe Gen5 and Infinity Fabric signal vias on the UBB require controlled-depth backdrilling to remove through-hole via stubs. At the UBB's typical board thickness (3.5–5 mm for 16–24 layers), stub lengths for signals connecting to intermediate layers can be 1–3 mm without backdrilling—sufficient to create resonances within the PCIe Gen5 signal band at 16 GHz. Backdrill depth accuracy of ± 50 μm is required; UBB-specific backdrill depth files are generated from the measured as-built stackup thickness of each production panel, as detailed in the PCIe Gen5 PCB design guide.
OAM slot connector footprints: The UBB carries 8 OAM slot connector receptacles (one per module position). These connectors are typically press-fit or surface-mount SMT types rated for the OAM Gen2 electrical specification (PCIe Gen5, 48 V power, Infinity Fabric). The connector footprint locations on the UBB must be held to ± 0.1 mm positional tolerance to ensure consistent module engagement geometry across all 8 slots. Connector coplanarity (the variation in Z-height among all 8 connector mating surfaces) must be < 0.05 mm to ensure that all 8 modules engage simultaneously when the UBB is installed in the chassis.
SMT assembly of OAM modules involves placing and reflowing several categories of components with very different thermal mass, package size, and solder joint quality requirements simultaneously. The assembly sequence is designed to manage these competing requirements.
Solder paste printing: A laser-cut stainless steel stencil with apertures matching the OAM module pad layout is aligned to the PCB fiducials and paste is squeegeed through the apertures. Stencil thickness is typically 100–120 μm for standard SMT pads; step stencil technology (locally reduced stencil thickness over fine-pitch BGA areas) is used when the accelerator package BGA pitch is below 0.65 mm and standard stencil thickness would produce excess paste volume. 3D solder paste inspection (SPI) immediately after printing measures paste volume, height, and offset on every pad; paste volume deviation exceeding ± 20% on accelerator or HBM BGA pads is cause for board rejection and restenciling before placement.
Accelerator package placement: The AI accelerator package (for example, AMD MI300X in OAM form factor, which uses TSMC's 3D chiplet packaging with multiple XCD dies and HBM stacks on a shared interposer) is the largest, heaviest, and most thermally massive component on the module. It is typically supplied in a tray and placed by a gantry-type SMT machine with vision-based fiducial alignment. Placement accuracy of ± 25 μm is required to ensure that all BGA balls are centered on their corresponding pads; placement error beyond 30–40% of ball pitch creates solder bridging or open joints that will not be detected until X-ray or functional test.
Supporting component placement: Bulk decoupling capacitors (0402 and 0201 sizes), VRM components (power inductors, gate driver ICs, MOSFET arrays), and management ICs are placed before or after the accelerator package depending on the assembly sequence. On-module VRMs must be placed with attention to the inductors' orientation relative to the PCB thermal vias beneath them; inductors that bridge a thermal via array and a nearby signal connector create assembly-specific thermal gradients that affect long-term solder joint reliability.
Reflow soldering: The reflow profile for an OAM module is dominated by the thermal mass of the accelerator package. An AMD MI300X-class package—a multi-die chiplet assembly approximately 60 mm × 60 mm with four HBM stacks—has substantial thermal mass that slows temperature rise during the preheat and soak zones of the reflow profile. Key profile parameters:
Package warpage during reflow is the primary risk for head-on-pillow (HoP) defects on large accelerator packages. HoP occurs when the PCB solder paste dome and the BGA ball do not fully coalesce during reflow because the package lifts slightly from the board during the liquidus phase as CTE-driven warpage temporarily separates ball from paste. Nitrogen atmosphere, precise soak timing, and shadow-moiré or simulation-based warpage prediction for the specific package are the primary mitigation tools.
The OAM edge connector is not a separate component soldered to the PCB—it is formed by the PCB edge fingers themselves, plated with electrolytic nickel and hard gold. This design simplifies the module assembly (no connector placement step) but places the quality of the PCB edge finger plating as a critical path item for the entire module's electrical reliability.
The edge finger plating sequence on the OAM module PCB is:
Gold plating quality is verified at three points: XRF thickness measurement (minimum 5 measurement points distributed across the finger array, including both ends and center); visual inspection under 10× magnification for scratches, pinholes, or bare copper areas; and contact resistance measurement across representative finger pairs to verify that the plated contact resistance is below the OAM specification limit (< 10 mΩ per contact at rated insertion force).
Common edge connector failure modes are: insufficient gold thickness at the board corners where plating bath agitation is lowest (corrected by extended plating time or bath parameter adjustment); gold plating contamination from handling after plating (corrected by packaging edge fingers immediately after plating in anti-static bags); and micro-cracks in the nickel layer at the finger tip during bevel routing (corrected by sharp routing tools and reduced cutting speed near the plated surface).
The OAM specification requires that the module's thermal contact surface—the top surface of the accelerator package or a metal heat spreader bonded to it—be flat to within < 0.1 mm across the contact area. This flatness tolerance is necessary to ensure that the cold plate or heatsink in the chassis achieves full-area contact across the thermal interface material (TIM) layer, which in turn is necessary to maintain junction temperature within specification at 700–750 W continuous dissipation.
After SMT reflow, warpage in the assembled module can cause the top surface of the accelerator package to deviate from the OAM flatness specification in several ways. The package itself may have bow from the reflow thermal cycle (CTE mismatch between the PCB laminate and the package body). The PCB may have residual warp from the reflow profile or from asymmetric copper distribution in the stackup. The combination of package and PCB warp determines the actual thermal interface surface profile of the assembled module.
Flatness measurement is performed using a laser profilometer or white-light interferometer that maps the full thermal contact surface at a grid of measurement points. Modules with surface deviation exceeding 0.1 mm across the contact area are either rejected or subjected to a controlled rework process (targeted thermal cycle to reduce warpage) before re-measurement. The measurement data is retained as part of the module's production record.
For modules where the thermal interface surface is a separate metal lid (copper or aluminum alloy) bonded to the package top with a die attach film or TIM2 material, the lid flatness and bonding quality must also be verified. Air gaps beneath the lid caused by inadequate TIM2 coverage create localized hot spots that may not be detectable without infrared thermography during functional test at full power.
OAM module inspection is more extensive than standard PCB assembly inspection because of the high component value and the consequences of a latent defect reaching a data center deployment.
Automated Optical Inspection (AOI): Post-reflow AOI inspects all SMT components for correct placement, polarity, and solder fillet quality on accessible joints. AOI cannot inspect the underside of the accelerator package or HBM package BGA arrays, which are hidden by the package body. AOI detects: missing or wrong-value passive components; lifted or skewed components; solder bridges on accessible fine-pitch connections; and tombstoned components on small passives.
3D X-Ray Inspection (AXI): Every OAM module undergoes 3D computed tomography X-ray inspection of the accelerator package and HBM package BGA arrays. 3D AXI reconstructs a volumetric image of the solder ball array from multiple 2D projections at different angles, revealing internal voids, partial bridges, and HoP defects that 2D X-ray cannot detect. Acceptance criteria per IPC-7095: void area < 25% of ball cross-section for any single ball; no voids in more than 10% of balls in the array; no bridges between adjacent pads. Solder balls that fail these criteria are flagged for disposition: modules with isolated single failing balls in non-critical signal regions may be accepted under engineering review; modules with failing balls in power or high-speed signal regions are rejected.
Edge connector dimensional inspection: The OAM edge connector finger array is measured for coplanarity (the variation in Z-height of all fingers relative to a datum plane) using a contact profilometer or optical measurement system. Coplanarity must be < 0.1 mm across the full finger array to ensure simultaneous contact with the UBB receptacle. Individual finger width and pitch are measured against the OAM connector specification to verify that the PCB routing and etching process has not introduced systematic feature size errors.
Thermal interface flatness verification: As described in the preceding section, laser profilometry of the thermal contact surface is a mandatory inspection step. The measurement report is part of the module's traceability record.
Dimensional and mechanical inspection: Module outline dimensions, mounting hole locations (if present), edge connector bevel geometry, and module height (clearance above the PCB top surface) are verified against the OAM mechanical specification. Modules exceeding dimensional tolerances in any of these parameters will not mate correctly with a standard UBB or chassis cold plate.
Functional test verifies that all electrical subsystems of the OAM module operate correctly before the module is shipped. For AI accelerator OAM modules, the functional test sequence typically includes:
Power-on and rail sequencing test: The module is connected to a test UBB or dedicated test fixture that provides the OAM-specified power rails (48 V bus) in the correct power-on sequence. Each power rail is verified to reach the correct voltage within the specified power-on timing. Current draw on each rail is monitored; excessive current (overcurrent) indicates a short circuit or mis-loaded component. Absent rails indicate an open circuit in the power delivery chain.
PCIe Gen5 enumeration test: The test fixture connects the module's PCIe Gen5 ×16 interface to a host system (or PCIe test controller). Successful enumeration at Gen5 speed (32 GT/s) with all 16 lanes active confirms that the PCIe signal integrity is adequate and that the accelerator package's PCIe PHY is functional. Link training failures at Gen5 that succeed at Gen4 indicate a marginal PCIe Gen5 channel, which may be caused by inadequate board material, via stub resonance, or solder joint defects on the accelerator BGA.
Memory bandwidth test: A standardized memory bandwidth benchmark (for example, a streaming read/write test) verifies that all HBM stacks are functional and achieving the specified bandwidth. AMD MI300X-class modules should achieve their rated 5.3 TB/s aggregate HBM3 bandwidth within measurement tolerance. A result significantly below specification indicates HBM stack connectivity issues (bridged or open BGA joints between the accelerator package interposer and the HBM package) or insufficient voltage on the HBM power rail.
Compute functional test: A targeted compute workload (matrix multiplication benchmark or equivalent) verifies that the AI compute cores are functional and achieving expected throughput. This test is less sensitive to subtle solder joint defects than the memory bandwidth test but catches gross die defects and power delivery failures.
Burn-in: Modules are operated at elevated ambient temperature (55–70°C) at maximum compute load for 24–72 hours. Burn-in screens for infant mortality failures—latent defects in solder joints, via plating, or die connections that manifest early in operation. Modules that fail during burn-in are removed and analyzed; failure analysis data is fed back into the assembly process to identify and correct systematic yield detractors. Modules that survive burn-in with all functional parameters within specification are released to shipping.
Thermal imaging during burn-in: Infrared (IR) camera imaging of the module surface during burn-in at full power identifies hot spots that indicate high-resistance electrical connections, missing thermal interface material, or cold plate contact gaps. Hot spots above a threshold (typically 15–20°C above the neighboring module surface at equivalent power dissipation) trigger removal for physical inspection and rework.
OAM modules require specialized packaging to protect the edge connector fingers, thermal interface surface, and fine-pitch BGA package areas during transit and storage.
The standard OAM module packaging approach uses: an anti-static foam insert molded to the module outline that supports the PCB and component side without contact pressure on the edge connector fingers or the thermal interface surface; a thermoformed plastic tray that holds 1–4 modules with positive mechanical retention; a moisture barrier bag (MBB) with desiccant and a humidity indicator card, heat-sealed after insertion; and an outer corrugated shipping carton with shock-absorbing foam padding rated to the ISTA transport test profiles applicable to the shipping distance and mode.
The edge connector fingers are protected by a removable plastic finger guard during packing. The finger guard covers the full finger array and is retained by friction clips at the board edges; it is removed immediately before UBB installation. Modules must not be handled by the edge connector area after finger guard removal; handling must be by the PCB substrate edges or by the mechanical mounting features if present.
The moisture sensitivity level (MSL) of OAM modules is determined by the most moisture-sensitive component on the board. AI accelerator packages with large, multi-die CoWoS or 3D chiplet constructions are typically MSL 3 or MSL 2a (requiring baking before use if MBB seal time exceeds the floor life specification). The MBB label must state the MSL rating and the cumulative floor life remaining at the time of packing. Data centers receiving OAM modules from extended storage should verify MBB integrity before opening and bake modules that have exceeded floor life according to IPC/JEDEC J-STD-033.
| Failure Mode | Root Cause | Detection Method | Prevention |
|---|---|---|---|
| Head-on-pillow (HoP) solder joint on accelerator BGA | Package warpage during reflow separates ball from paste before coalescence | 3D X-ray AXI; functional test failure (intermittent) | Optimized reflow profile; nitrogen atmosphere; package warpage simulation |
| Solder void > 25% on power BGA balls | Flux outgassing trapped in via-in-pad; solder paste contamination | 3D X-ray AXI | Via-in-pad epoxy fill planarization within spec; controlled paste storage; nitrogen reflow |
| PCIe Gen5 link training failure | Via stub resonance; insufficient laminate quality; open signal ball | Functional test (PCIe enumeration) | Backdrill to < 10 mil stub; low-loss laminate on PCIe layers; 3D X-ray on signal ball rows |
| Edge connector contact resistance above limit | Insufficient gold thickness; handling damage to fingers; plating contamination | Contact resistance measurement; XRF gold thickness | Verified plating process; immediate post-plating packaging; handle by PCB body only |
| Thermal interface flatness out of spec | Package and/or PCB warpage after reflow | Laser profilometry | Controlled reflow cool-down rate; symmetric copper distribution in stackup; warpage measurement at multiple production stages |
| HBM bandwidth below specification | Open or bridged HBM-to-interposer BGA balls within CoWoS package | Memory bandwidth functional test | Optimized reflow for CoWoS package mass; 3D X-ray at HBM ball array; verify HBM power rails |
| VRM thermal runaway during burn-in | Insufficient thermal via density under VRM components; incorrect inductor placement over via keepout | IR thermography during burn-in | Verify thermal via array density per design specification; DFM review of VRM component placement vs thermal keepout zones |
| Infant mortality failure within 24 hours of field deployment | Latent solder joint crack from thermal cycling; marginal via plating | Burn-in; IR thermography | Adequate burn-in duration (48–72 hours at maximum load); IPC Class 3 via plating inspection |
A common question from engineering teams transitioning from NVIDIA SXM-based H100 infrastructure to OAM-based MI300X infrastructure is how the manufacturing process differs. The answer is that OAM module manufacturing and SXM-based baseboard manufacturing share many of the same process steps but differ in how those steps are distributed between the module and the baseboard. For context on the SXM ecosystem and its PCB implications, see A100 vs H100: PCB Stack Differences and the Blackwell architecture guide.
| Manufacturing Dimension | OAM Module + UBB | NVIDIA SXM Baseboard |
|---|---|---|
| Accelerator package assembly | On OAM module PCB (small, dedicated board) | On large baseboard alongside NVSwitch, VRMs |
| Module PCB complexity | Medium (12–20 layers; compact 74 mm × 74 mm) | N/A (GPU is on the baseboard itself) |
| Baseboard / UBB PCB complexity | High (16–22 layers; large format; PCIe Gen5 + Infinity Fabric routing) | Very high (20–32 layers; NVLink 4.0/5.0; NVSwitch on board) |
| Edge connector | Hard gold plated PCB fingers on OAM module; receptacle on UBB | LGA socket on baseboard; no separate module connector |
| Accelerator package rework | Module-level rework (remove and replace OAM module); UBB is not affected | Board-level rework (GPU package rework directly on large baseboard; high risk) |
| Yield management | Module yield and UBB yield are independent; a failing module does not scrap the UBB | A failing GPU package on the baseboard may require scrapping the entire board |
| Thermal interface | Defined at OAM module top surface; must meet flatness spec independently | Cold plate contacts GPU package through SXM socket; baseboard does not define the thermal interface |
| Inter-GPU interconnect routing | Infinity Fabric routed on UBB between OAM slots; no switch chip on UBB | NVLink routed on baseboard between GPU packages and NVSwitch chips (very high routing density) |
| Supply chain flexibility | Any OAM-compliant module from any vendor in the same UBB slots | Only NVIDIA SXM GPUs in NVIDIA SXM sockets |
The OAM architecture's separation of the accelerator module from the baseboard (UBB) has a practical manufacturing advantage: module-level yield loss does not impact the UBB. A failing MI300X OAM module is replaced with a new module; the UBB, which costs significantly more to fabricate and assemble than a single OAM module, is not affected. In the SXM architecture, a failing GPU package on a large baseboard triggers a difficult and risk-prone board-level rework operation or scraps the entire assembly.
What IPC class standard applies to OAM module PCBs?
IPC Class 3 (High Reliability) applies to OAM module PCBs and UBBs. Class 3 specifies the most stringent requirements for via barrel copper thickness, annular ring, hole quality, and surface finish coverage. Given that OAM modules operate continuously at 700–750 W in data center environments for multi-year service lives, Class 3 fabrication combined with extended burn-in testing is the minimum responsible specification. The same Class 3 requirements apply to UBBs, which carry high-current 48 V power planes and must maintain electrical integrity through tens of thousands of thermal cycles over their operating life.
How many insertion/removal cycles must OAM edge connectors support?
The OAM specification defines the minimum edge connector insertion cycle life based on the OAM module use case. For standard data center deployment where modules are inserted during system assembly and replaced only on failure, the minimum specification is typically 100 insertion cycles. For hot-swap or high-turnover deployments, the specification may require 200+ cycles. The gold plating thickness specification (0.76–1.27 μm minimum) is set to ensure that adequate gold remains on the contact surface after the maximum specified insertion cycles; thinner gold wears through to the nickel underlayer, raising contact resistance and risking copper corrosion at the nickel/copper interface.
Can OAM modules be reworked in the field if an accelerator package fails?
OAM modules are designed as field-replaceable units (FRUs), meaning that a failed module is replaced at the module level rather than repaired. The module slots out of the UBB edge connector, and a replacement module slots in; the UBB and chassis are not disturbed. Board-level rework of the accelerator package on an OAM module PCB (removing and replacing the AI accelerator BGA) is technically possible but is generally performed only by the original manufacturer or a specialized rework service, not in the field. The risk of further damage to the high-value accelerator package substrate during rework typically makes module replacement the preferred option for field failures.
What is the typical production lead time for OAM module prototypes?
OAM module prototype production lead time depends on PCB complexity and component availability. For a standard OAM module PCB with 2+N+2 HDI (three lamination cycles), via-in-pad processing, and edge connector gold plating, bare board fabrication takes approximately 15–20 business days. SMT assembly with accelerator package placement, reflow, X-ray inspection, functional test, and burn-in adds approximately 10–15 business days. Total prototype lead time from design file release to shipped, tested module is typically 5–7 weeks. Production volumes with established supply chains and qualified processes can achieve 3–4 week lead times.
What is the moisture sensitivity level (MSL) for OAM modules with AMD MI300X?
MSL ratings for specific OAM modules depend on the accelerator package used. AMD MI300X in OAM form factor uses a multi-die chiplet package that, like other large advanced packaging constructions, is typically rated MSL 3. MSL 3 means the module can be exposed to ambient conditions (30°C / 60% RH or less) for a cumulative 168 hours (7 days) after MBB opening before solder reflow. Modules stored beyond this floor life must be baked per J-STD-033 before use. The receiving data center or system integrator should verify that the MBB was sealed at time of manufacture and that the floor life has not been exceeded before opening the package.
Does the UBB need to be replaced when upgrading from one OAM GPU generation to the next?
This depends on whether the new module generation is electrically and mechanically compatible with the existing UBB. OAM Gen1 and Gen2 maintain mechanical connector compatibility, meaning a Gen2 module can physically plug into a Gen1 UBB slot. However, if the Gen2 module requires 48 V power delivery and the Gen1 UBB only provides 12 V, or if the Gen2 module uses Infinity Fabric at higher bandwidth than the UBB routing was designed for, the UBB will need to be redesigned or replaced to fully support the new module. As a practical matter, UBBs are often designed with the specific module generation in mind; organizations planning a generational GPU upgrade should verify UBB compatibility before assuming the existing UBB infrastructure can be reused.
>> Recommend reading: Ultimate PCB Assembly Guide: Prototyping to Mass Production | NextPCB
NextPCB supports the complete OAM module and UBB manufacturing process: HDI PCB fabrication with via-in-pad, edge connector gold plating, SMT assembly with large-format BGA placement, 3D X-ray inspection, functional test, and burn-in. Our capabilities cover the full range of AI accelerator assembly requirements from prototype to production volum
Still, need help? Contact Us: support@nextpcb.com
Need a PCB or PCBA quote? Quote now
