High Current GPU Power Delivery PCB
Quality & Inspection Standards
PRODUCTS CENTER
Get Custom Quote
PRODUCTS CENTER
High Current GPU Power Delivery PCB
Delivering 600 W to a GPU through a PCB is not a simple copper thickness problem. At core voltages below 1 V, a 5% IR drop is only 50 mV — lost in the board before the silicon sees it. The design must simultaneously manage DC resistance across heavy copper planes, via current density under the GPU footprint, resin fill between thick layers, and thermal gradients that shift the operating point. This page defines the engineering thresholds where a GPU power delivery design moves from functional to reliable.
Engineering Decision Map
1. The Power Delivery Failure Chain
GPU power delivery failures follow a predictable chain that begins at the PCB design level and ends at the transistor:
IR drop chain: Undersized power planes or insufficient via count → IR drop across the PCB exceeds 50 mV → GPU core voltage dips below minimum at peak current → timing violations or logic errors → system crash under load.
Thermal chain: High current density in localized via clusters → Joule heating → temperature rise above 20 °C in the via region → increased copper resistivity (0.4%/°C) → further IR drop → thermal runaway in the hottest vias → barrel failure or pad cratering.
Delamination chain: Resin starvation between 4 oz copper planes → void formation during lamination → trapped moisture expansion during reflow → internal delamination → increased thermal resistance → hot spot → plane‑to‑plane short.
2. Copper Weight and Resin Fill: The Lamination Constraint
Carrying 150 A across a GPU power plane requires copper weights of 3‑6 oz. The electrical requirement is straightforward — the manufacturing requirement is not. Heavy copper planes must be separated by prepreg with sufficient resin to fill the gaps between conductors and bond to the copper surface.
Resin starvation threshold: Per IPC‑6012E §3.6.2.18, the pressed prepreg thickness must be ≥ 1.5× the copper thickness it covers. For 4 oz copper (140 µm), the minimum pressed dielectric thickness is 210 µm — typically requiring two plies of 2116 or one ply of 7628 with resin content ≥ 55%. Below this threshold, voids form during lamination and become delamination initiation sites.
| Copper Weight | Thickness (µm) | Min Pressed Prepreg (µm) | Recommended Glass Style | Resin Content Requirement |
|---|---|---|---|---|
| 2 oz | 70 | 105 | 1×1080 | ≥ 50% |
| 3 oz | 105 | 158 | 2×1080 or 1×2113 | ≥ 55% |
| 4 oz | 140 | 210 | 2×2116 or 1×7628 | ≥ 55% |
| 6 oz | 210 | 315 | 2×7628 | ≥ 55% |
Manufacturing reality: Beyond 4 oz, the prepreg must be matched to the copper thickness on both sides of the interface. A 4 oz layer adjacent to a 1 oz layer requires enough resin to fill the 4 oz side while maintaining minimum dielectric spacing on the 1 oz side — a condition that often forces multi‑ply prepreg with a high‑resin‑content center ply.
3. Via Current Density and Via Farm Design
The GPU power path transitions from the plane to the BGA pads through a via farm. Each via carries a fraction of the total current. Undersizing the via farm or clustering vias too tightly creates localized hot spots that degrade reliability.
Via current capacity per IPC‑2152: A 0.25 mm plated through‑hole with 25 µm barrel copper carries approximately 1.5 A at a 10 °C temperature rise. For a GPU rail carrying 150 A, the minimum via count is 100. A 10 °C rise is the design target; exceeding 20 °C accelerates barrel fatigue and CAF growth.
| Via Diameter (mm) | Copper Thickness (µm) | Current at 10 °C Rise (A) | Vias Needed for 150 A |
|---|---|---|---|
| 0.20 | 25 | 1.2 | 125 |
| 0.25 | 25 | 1.5 | 100 |
| 0.30 | 30 | 2.0 | 75 |
| 0.35 | 35 | 2.5 | 60 |
Via farm distribution: Clustering all 100 vias in a 10 × 10 mm area under the GPU die creates a local hot spot where the board temperature can exceed the ambient by 30 °C. The via farm must be distributed across the entire GPU package footprint, with spacing ≥ 1.0 mm between vias to allow lateral heat spreading.
Via‑in‑pad requirement: Under the GPU BGA, vias must be filled with non‑conductive epoxy and capped with planar copper per IPC‑6012E §3.3.4. An unfilled via under a BGA pad causes solder wicking during reflow, creating an open joint. The cap thickness must be ≥ 12 µm for Class 3, and the dimple must not exceed 25 µm.
4. Thermal Management and Hot Spot Prevention
The GPU itself dissipates 500‑700 W, and the PCB power delivery contributes an additional 20‑50 W of I²R heating in the copper planes and vias. The thermal design must prevent the board temperature from exceeding the material Tg during sustained operation.
Temperature rise in inner power planes: A 4 oz copper plane carrying 150 A across a 200 mm width generates approximately 0.5 W/cm² of heat. In still air, this can raise the internal board temperature by 15‑25 °C above ambient. Forced convection or conduction to a heatsink through thermal vias is required.
Thermal via requirements: Thermal vias under the GPU must connect the top‑side copper pad to the inner power planes and the bottom‑side ground plane. A thermal via array with 0.30 mm diameter, 1.0 mm pitch, and copper fill provides approximately 50 W/m·K effective through‑plane thermal conductivity. The array must cover the entire GPU die shadow area.
5. Power Plane Stackup and Voltage Drop Budget
The total IR drop from the voltage regulator to the GPU BGA pads must be ≤ 5% of the core voltage. For a 0.8 V GPU core, that is 40 mV. This budget must be allocated across the PCB power planes, vias, and connectors.
Plane resistance calculation: For a 4 oz copper plane (140 µm thick), the sheet resistance is approximately 0.12 mΩ per square. A 200 mm × 200 mm plane segment has roughly 1 square of resistance, contributing 0.12 mΩ. At 150 A, the IR drop across this segment is 18 mV — already consuming nearly half the budget.
Stackup strategy: GPU power planes must be placed close to the surface to minimize via length and IR drop. A typical 26‑layer AI server stackup places the GPU Vcore plane at L3‑L4 and L23‑L24, with multiple parallel planes to reduce effective resistance. The planes must be stitched with vias every 25 mm to maintain voltage uniformity across the plane.
Copper balance requirement: The heavy copper power planes on one side of the stackup must be mirrored by equivalent copper weight on the opposing side, even if it is a dummy fill. Asymmetry exceeding 5% copper density triggers warpage excursions above 0.5% on panels > 500 mm.
High Current GPU Power Delivery Checklist
- Specify copper weight ≥ 3 oz for GPU power planes carrying > 100 A. 4 oz is typical for 150 A designs.
- Ensure prepreg resin content ≥ 55% between heavy copper planes. Pressed thickness must be ≥ 1.5× copper thickness per IPC‑6012E §3.6.2.18.
- Calculate via count using IPC‑2152 with a 10 °C temperature rise target. Minimum 100 vias for 150 A at 0.25 mm diameter with 25 µm copper.
- Distribute vias evenly across the GPU package footprint with ≥ 1.0 mm spacing. Avoid clustering to prevent hot spots.
- Use filled and capped via‑in‑pad under BGA pads per IPC‑6012E §3.3.4. Cap thickness ≥ 12 µm for Class 3; dimple ≤ 25 µm.
- Place GPU power planes close to the surface to minimize via IR drop. Stitch parallel planes with vias every 25 mm.
- Mirror heavy copper planes on opposing sides to maintain stackup symmetry. Copper density mismatch must not exceed 5%.
- Validate with thermal imaging and IR drop measurement on first‑article boards at full load. Hot spots must not exceed 20 °C above ambient.
FAQ – High Current GPU Power Delivery PCB
Q: How much copper thickness is enough for a 150 A GPU rail?
A: 4 oz copper with a 200 mm plane width is sufficient for most designs, provided the via farm is adequately sized. If the plane must carry current more than 250 mm, consider 5‑6 oz or split the current across two parallel planes.
Q: What happens if vias are clustered under the GPU die?
A: Clustered vias create a localized hot spot where the board temperature can exceed the ambient by 30 °C or more. This accelerates barrel fatigue and CAF growth, and can cause the board to exceed its Tg in that region.
Q: Can I use unfilled vias under a GPU BGA?
A: No. Unfilled vias under BGA pads cause solder wicking during reflow, which creates open joints. Filled and capped vias per IPC‑6012E §3.3.4 are mandatory for Class 2 and Class 3 designs.
Q: How does heavy copper affect PCB warpage?
A: Heavy copper planes create asymmetric internal stress if not mirrored on the opposing side of the stackup. The copper density mismatch between opposing halves must be ≤ 5% to keep warpage below 0.5% on panels exceeding 500 mm.
Q: What is the most common GPU power delivery failure?
A: Via barrel fatigue caused by localized overheating in a clustered via farm. The hottest vias fail first, shifting current to the remaining vias and accelerating the failure cascade.
Related Engineering Pages
Warpage control and 224G stackup rules. Low Loss Material Selection for AI PCB →
Dk/Df and copper roughness for high‑speed layers. Thermal Reliability in AI Server PCB →
Managing sustained high‑temperature operation. 112G / 224G PAM4 Backplane PCB →
Signal integrity rules for backplane traces.
Need a GPU power delivery review, via farm analysis, or heavy copper stackup validation?
Request Engineering Review →References: IPC‑6012E §3.6.2.18 (Dielectric Spacing), §3.3.4 (Via‑in‑Pad), IPC‑2152 (Current Carrying Capacity), IPC‑TM‑650 (Thermal Stress). Courtesy of IPC.
Get Custom Quote
PRODUCTS CENTER


