The SXM (Socketed Mezzanine Module) is NVIDIA's proprietary GPU form factor for HGX platforms — a high-density mezzanine card that carries the GPU die, HBM stacks, and the NVLink interconnect fabric. Unlike PCIe add-in cards, SXM modules connect through a high-pin-count mezzanine socket directly to the HGX baseboard, enabling dramatically higher bandwidth. This article unpacks the PCB design challenges of SXM carrier boards.

SXM Generations: Evolution of Bandwidth

Each generation doubles NVLink bandwidth while increasing TDP — the SXM6 at 1,000W represents a staggering power density of 5.7 W/cm² across the module area.

SXM Connector: The Critical Interface

The SXM mezzanine connector is a proprietary NVIDIA-designed interface with approximately 3,000–3,500 pins. Key characteristics:

• Signal pins: ~2,000 pins dedicated to NVLink high-speed differential pairs (18 ports × 4 lanes × 2 pairs/lane = 144 pairs per direction)

• Power pins: ~1,000 pins for power delivery — each pin rated at 1.5A continuous. Total connector current capacity exceeds 1,500A for the combined voltage rails

• Contact resistance: <15 mΩ per pin after mating, including connector body resistance

• Mating height: ~5 mm stack height — the SXM module sits 5 mm above the baseboard

• Alignment tolerance: ±0.15 mm in X/Y; the connector includes guide pins and a metal stiffener frame for precision alignment

PCB Stackup: 20-28 Layer Design

An SXM5 carrier PCB typically uses 22-26 layers:

• Top layer: GPU BGA (0.4 mm pitch, ~55 mm × 55 mm), HBM3 stacks (×6), connector placement zone

• Signal layers (12–16): NVLink routing to connector, HBM PHY breakout, management signals

• Power planes (6–8): Vcore (0.75–0.85V at 600–800A), Vmem-HBM (1.1V at 60–80A), Vddq (1.2V at 30A)

• Material: Megtron 7 for all signal layers (NVLink at 100 GB/s PAM4 demands ultra-low loss); high-Tg FR-4 for power layers only

NVLink Signal Integrity at 100 GB/s PAM4

NVLink 4 (Blackwell) operates at 100 GB/s per lane using PAM4 signaling at 50 Gbaud. The Nyquist frequency is 25 GHz. Critical SI requirements:

• Insertion loss: Must stay below −25 dB at 25 GHz across the complete channel (GPU die → SXM PCB → SXM connector → baseboard PCB → NVSwitch). SXM PCB allocation is ~6 dB

• Via design: All NVLink vias must be backdrilled to within 6 mil of the target layer. Via stub resonance at 25 GHz (λ/4 ≈ 2.5 mm in Megtron 7) is catastrophic if not removed

• Pair-to-pair skew: <1 ps within each differential pair. Trace length matching within 3 mil

• Impedance: 100Ω differential ±7%

• Return loss: < −15 dB across 10 MHz–25 GHz — the connector is typically the dominant discontinuity

Power Delivery at 1,000W

The SXM6's 1,000W TDP pushes power delivery to extremes:

• GPU core current: At 0.75V Vcore, 1,000W means 1,333A — though approximately 30% is HBM power, making actual Vcore current ~900A

• VRM placement: Multi-phase VRM on the baseboard (not the SXM module). The SXM connector carries 0.8V core voltage directly — the PCB just distributes it to the GPU BGA

• IR drop: Total path resistance from connector pins to GPU die bumps must be below 50 μΩ. At 900A, 50 μΩ = 45 mV drop — 5.6% of a 0.8V rail

• Decoupling: Thousands of MLCCs on the SXM PCB in a dense grid under the GPU — reverse-geometry 0306 capacitors with <0.15 nH mounted inductance

• Plane count: 6–8 copper planes at 3–4 oz weight dedicated to Vcore distribution

Thermal and Mechanical Design

SXM modules are the hottest components in the server, requiring direct liquid cooling or massive vapor chamber heatsinks:

• Heatsink mounting: The heatsink bolts to the baseboard stiffener, clamping the SXM module between the heatsink and connector — not bolted to the SXM PCB itself

• Thermal interface: Phase-change TIM or liquid metal (e.g., Thermal Grizzly Conductonaut) between GPU lid and heatsink cold plate, achieving <0.05 °C·cm²/W thermal resistance

• PCB warpage: At 2.2 mm thickness with 0.4 mm pitch BGA, warpage must stay below 0.2% post-reflow. Asymmetric copper distribution causes severe warpage — stackup must be carefully balanced

• Underfill: Capillary underfill under the GPU BGA for mechanical reliability; must be compatible with the high-temperature liquid cooling environment (85°C continuous)