Surface
Julia Wu - Senior Sales Engineer at NextPCB.com
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
A standard desktop motherboard uses 6–8 PCB layers. A high-end enterprise server board uses 10–14. A DGX H100 GPU baseboard uses 20–24. A dedicated NVSwitch board in a GB200 NVL72 rack uses 32–40 or more. This progression is not incremental—it reflects a fundamental change in the engineering demands that AI accelerator hardware places on printed circuit boards.
The question “why do AI GPU PCBs need so many layers?” has a precise engineering answer, and understanding it matters for anyone designing, specifying, or procuring PCBs for AI server infrastructure. Layer count is not a measure of quality or prestige—it is the outcome of four specific engineering requirements that converge in AI GPU baseboards in a way that they do not in any other class of commercial electronics. This article explains each of those requirements in detail.
PCB layer count is determined by the need to route electrical connections between components in a way that satisfies all applicable constraints simultaneously. Those constraints are:
In a standard server motherboard, these constraints lead to 10–14 layers. In an AI GPU baseboard, all of these constraints are pushed to their limits simultaneously—leading to layer counts that would have been considered impractical in commercial electronics a decade ago.
The dominant driver of high layer count in AI GPU baseboards is the sheer number of high-speed differential pairs that must be routed between GPU packages and NVSwitch packages.
In a DGX H100 baseboard with 8 H100 GPUs and 4 NVSwitch 3.0 chips:
Each of these differential pairs must be routed as a controlled-impedance structure (100 Ω ± 5%) with:
At 100 Gb/s per lane (NVLink 4.0), these constraints cannot be relaxed. A 75 μm wide trace with 75 μm space between pairs (3 mil / 3 mil) can route approximately 8–10 differential pairs across a 100 mm wide routing channel per layer. Routing 2,000+ pairs across the board therefore requires a minimum of 6–10 dedicated NVLink signal routing layers, depending on board width and routing efficiency.
For B200 boards with NVLink 5.0 at 200 Gb/s per lane, the spacing requirements are tighter (lower crosstalk budget), and the number of NVLink links per GPU increases. NVSwitch 4.0 boards in the GB200 NVL72, routing NVLink 5.0 connections between 72 GPUs via 9 switch boards, push signal routing density to the extreme end of what current PCB manufacturing can achieve.
AI GPU baseboards manage many distinct power domains simultaneously. Each domain that requires a dedicated power plane (rather than sharing a plane with other domains) adds layers to the stackup. For a DGX H100 baseboard, the major power domains include:
| Power Domain | Voltage | Current (per 8-GPU board) | Plane Requirement |
|---|---|---|---|
| GPU VCORE (per GPU) | ~0.85–0.9 V | ~500 A per GPU × 8 = 4,000 A total | Dedicated planes per GPU; high copper weight |
| GPU HBM VDDQ | ~1.1 V | ~50–100 A per GPU | Shared or dedicated; noise-sensitive |
| NVSwitch VCORE (per chip) | ~0.8 V | ~200–300 A per chip × 4 chips | Dedicated planes for each NVSwitch |
| NVLink I/O voltage | ~1.0–1.1 V | High aggregate current | Separate from VCORE; noise isolation required |
| PCIe I/O voltage | ~0.85–1.0 V | Moderate | Can share with other I/O rails if noise isolated |
| Management / BMC rails | 3.3 V, 1.8 V, 1.0 V | Low current | Can share layers; low priority for dedicated planes |
| Ground reference | 0 V | Return for all above | Multiple dedicated ground planes required |
The practical minimum for the power and ground plane layer count is 8–12 layers on a fully loaded H100 baseboard. Combined with the 6–10 NVLink signal layers and the HDI build-up layers, this accounts for the majority of the total 20–24 layer count.
For B200 baseboards with higher GPU TDP (1,000 W vs 700 W) and more power rails per GPU (dual-die B200 has more distinct power domains than the single-die H100), the power plane count increases further, pushing total layer count toward 28–32.
GPU packages (SXM5 for H100, SXM6 for B200) and NVSwitch chips are large, fine-pitch BGA components. Routing signals out of the inner rows of a BGA footprint requires vias that transition signals from the pad layer to inner routing layers—consuming layer transitions that are not available for other routing until the via escapes the BGA keepout zone.
For a 60 mm × 60 mm BGA package with 0.65 mm ball pitch, the inner rows of pads can only be accessed by routing signals through the outer rows using vias. Each via:
HDI via structures (laser-drilled microvias and via-in-pad) allow BGA escape to be distributed across multiple build-up layers, dramatically increasing the routing density that can be achieved beneath a fine-pitch BGA. Without HDI, a 64-row BGA package like NVSwitch 3.0 cannot be efficiently escaped in a board of practical thickness—the number of through-hole vias required would consume more routing channel space than the vias free up.
The HDI build-up layers required for BGA escape add 2–6 layers to each side of the core (1+N+1 to 3+N+3 configurations), directly increasing total layer count beyond what the core routing alone would require.
At NVLink 4.0 and NVLink 5.0 speeds, every signal routing layer must have an unbroken reference plane (ground) immediately adjacent to it. The characteristic impedance of a differential pair depends on the distance to the reference plane and the dielectric constant of the material between them. If the reference plane is interrupted (by a cutout, a connector clearance, or a power plane transition), the characteristic impedance changes locally, creating a reflection that degrades signal quality.
This means that signal layers and reference planes must alternate strictly: signal layer, reference layer, signal layer, reference layer. A board with 10 dedicated signal routing layers therefore requires a minimum of 10 reference (ground) plane layers interspersed among them—and this count does not include the power planes required by Driver 2.
In practice, ground planes often serve double duty (as both the reference for one signal layer above and the reference for a different signal layer below), reducing but not eliminating the requirement for dedicated ground layers. In a well-designed 28-layer stackup, approximately 8–10 layers are dedicated ground planes, with the remainder split between signal routing and power planes.
| Board Type | GPU Generation | Typical Layer Count | Primary Driver of High Count |
|---|---|---|---|
| Consumer GPU add-in card (RTX 4090) | Ada Lovelace | 8–12 | PCIe Gen4, GDDR6X routing, moderate PDN |
| Data center inference GPU (L40S PCIe) | Ada Lovelace | 12–16 | PCIe Gen4, GDDR6 routing, no NVLink |
| A100 HGX baseboard | Ampere | 14–18 | NVLink 3.0, 6 × NVSwitch 2.0, HBM2e power |
| H100 HGX baseboard | Hopper | 20–24 | NVLink 4.0, 4 × NVSwitch 3.0, PCIe Gen5 |
| MI300X OAM UBB | CDNA 3 | 16–22 | Infinity Fabric, PCIe Gen5, 48 V PDN |
| B200 GPU baseboard | Blackwell | 24–32 | NVLink 5.0, PCIe Gen6 PAM4, 1,000 W PDN |
| NVSwitch 3.0 board (H100 era) | Hopper | 28–36 | All NVLink 4.0 fabric routing; no GPU PDN |
| NVSwitch 4.0 board (GB200 NVL72) | Blackwell | 32–40+ | NVLink 5.0 at maximum routing density; 400 W per NVSwitch chip |
For detailed layer count analysis of specific GPU generations, see A100 vs H100: GPU Generational Leap & PCB Stack Differences Explained and NVIDIA Blackwell Architecture Explained.
HDI (High-Density Interconnect) refers to PCB technology that achieves higher routing density than standard through-hole via designs by using laser-drilled microvias, fine trace/space geometries, and sequential lamination build-up layers. HDI technology is not optional for AI GPU baseboards—it is a prerequisite for routing the BGA escape of GPU and NVSwitch packages within the board thickness and footprint constraints of a practical server form factor.
| HDI Type | Description | Layer Count Notation | Typical Use Case |
|---|---|---|---|
| Type I | One build-up layer on one or both sides; blind vias from outer layer to first inner layer; through-holes in core | 1+N+1 (both sides) or 1+N (one side) | Moderate-density BGA escape; data center server motherboards |
| Type II | Two build-up layers per side; stacked or staggered microvias; through-holes in core | 2+N+2 | H100 HGX baseboard; fine-pitch BGA escape for SXM5 and NVSwitch 3.0 |
| Type III | Three build-up layers per side; stacked microvias; through-holes in core | 3+N+3 | B200 baseboard; extreme BGA density; some NVSwitch boards |
| Any-Layer HDI (ELIC) | Every layer interconnectable via stacked filled microvias; no through-holes spanning full board thickness | All-layer microvia | GB200 NVL72 NVSwitch boards; maximum routing density at extreme layer counts |
Microvias are small-diameter (75–150 μm) vias drilled by laser rather than mechanical drill. They connect only two adjacent layers (blind vias) or span buried layers (buried vias), as opposed to through-hole vias that span the full board thickness.
Advantages for AI GPU PCBs:
Via-in-pad places a microvia (or filled through-hole via) directly within a BGA pad. The via is filled with conductive or non-conductive epoxy, planarized, and cap-plated to present a flat, solderable surface. This technique is essential for GPU and NVSwitch BGA escape because:
The manufacturing process for via-in-pad involves additional steps (epoxy fill, cure, planarization grinding, cap plating) and tighter process controls than standard via processing. See How GPU PCBs Are Manufactured: From Bare Board to Final PCBA for a detailed description of the via-in-pad manufacturing process.
Any-Layer HDI (also called ELIC—Every Layer Interconnect) takes HDI to its logical extreme: every layer in the stackup is interconnected by stacked, filled microvias. There are no traditional through-holes spanning the full board thickness. Signals can transition between any two layers using a stack of filled microvias, regardless of which layers are involved.
ELIC is the highest-complexity and highest-cost PCB structure in commercial production. It is reserved for applications where routing density requirements cannot be met with any other approach. In AI GPU hardware, ELIC is used on dedicated NVSwitch 4.0 boards in the GB200 NVL72, where the combination of 72 GPU NVLink 5.0 connections, 400 W per NVSwitch chip power delivery, and the physical constraints of the rack-integrated switch board format creates routing demands that exceed what 2+N+2 or 3+N+3 HDI can solve.
The following is a representative 28-layer stackup for a B200-class GPU baseboard with 2+N+2 HDI, illustrating how the layers are allocated among signal routing, power planes, and ground references:
| Layer | Function | Material | Copper Weight |
|---|---|---|---|
| L1 (outer) | SMT pads; BGA cap-plated via-in-pad; low-speed signals | Build-up prepreg (RCC) | 1 oz (ENIG finished) |
| L2 | Ground reference (HDI build-up layer 1) | Build-up prepreg | 1 oz |
| L3 | BGA escape routing; NVLink 5.0 breakout (build-up layer 2) | Build-up prepreg | ½ oz |
| L4 | Ground reference plane | Core laminate (Megtron 7) | 1 oz |
| L5 | NVLink 5.0 signal routing (horizontal) | Megtron 7 | ½ oz VLP copper |
| L6 | Ground reference plane | Megtron 7 | 1 oz |
| L7 | NVLink 5.0 signal routing (vertical) | Megtron 7 | ½ oz VLP copper |
| L8 | Ground reference plane | Megtron 7 | 1 oz |
| L9 | GPU VCORE power plane | Megtron 6 (or standard) | 2 oz |
| L10 | Ground plane (power return) | Megtron 6 | 2 oz |
| L11 | NVLink 5.0 signal routing (horizontal) | Megtron 7 | ½ oz VLP copper |
| L12 | Ground reference plane | Megtron 7 | 1 oz |
| L13 | NVLink 5.0 signal routing (vertical) | Megtron 7 | ½ oz VLP copper |
| L14 | PCIe Gen6 signal routing + NVLink 5.0 | Megtron 7 | ½ oz VLP copper |
| L15 | Ground reference plane (board center) | Megtron 6 | 1 oz |
| L16 | NVSwitch VCORE power plane | Megtron 6 | 2 oz |
| L17 | Ground plane (NVSwitch power return) | Megtron 6 | 2 oz |
| L18 | NVLink 5.0 signal routing (horizontal) | Megtron 7 | ½ oz VLP copper |
| L19 | Ground reference plane | Megtron 7 | 1 oz |
| L20 | NVLink 5.0 signal routing (vertical) | Megtron 7 | ½ oz VLP copper |
| L21 | HBM / auxiliary power planes (split plane) | Megtron 6 | 1 oz |
| L22 | Ground reference plane | Megtron 6 | 1 oz |
| L23 | NVLink 5.0 signal routing (horizontal) | Megtron 7 | ½ oz VLP copper |
| L24 | Ground reference plane | Megtron 7 | 1 oz |
| L25 | Management signals; low-speed I/O | Core laminate | 1 oz |
| L26 | BGA escape routing (build-up layer, bottom) | Build-up prepreg | ½ oz |
| L27 | Ground reference (HDI build-up layer, bottom) | Build-up prepreg | 1 oz |
| L28 (outer) | SMT pads; bottom-side component lands | Build-up prepreg (RCC) | 1 oz (ENIG finished) |
This representative stackup allocates: 8 NVLink 5.0 signal routing layers, 1 PCIe Gen6 layer, 2 BGA escape layers (HDI build-up), 10 ground/reference plane layers, 4 power plane layers (VCORE, NVSwitch, HBM), 1 management signal layer, and 2 outer pad layers. The Megtron 7 material is used on all NVLink and PCIe signal layers; Megtron 6 is used on power and ground planes where dielectric loss is irrelevant; build-up prepreg is used for the HDI layers.
Dedicated NVSwitch boards—such as those used in the GB200 NVL72 rack—represent the most demanding PCB design problem in current commercial production. These boards carry no GPU packages; their sole function is to switch NVLink 5.0 traffic between all 72 B200 GPUs in the rack.
This specialization makes them simultaneously simpler (no GPU power delivery requirements) and more demanding (higher NVLink routing density per unit board area) than GPU baseboards:
Layer counts of 32–40 or more are typical for NVSwitch 4.0 boards. These are among the most technically challenging PCBs in commercial production and are manufactured by fewer than a handful of qualified fabricators worldwide.
Higher layer count is not automatically better for signal integrity—it is a necessary accommodation for the routing constraints that arise when signal density is high. Several signal integrity effects are influenced directly by layer count decisions:
For a detailed treatment of signal integrity requirements at NVLink and PCIe Gen5/6 speeds, see What Is NVLink? How NVIDIA's High-Speed GPU Interconnect Shapes PCB Routing and the AI Accelerator PCB Design Guide.
Power plane layers in a high-layer-count board have lower sheet resistance (more copper area available for the same current) and lower inductance (more planes means more capacitance between adjacent power and ground planes, which provides high-frequency decoupling). Both effects improve PDN performance:
The relationship between layer count and thermal management is less direct than for signal integrity or PDN, but real:
A 28-layer 2+N+2 HDI board requires a minimum of three lamination press cycles:
Each press cycle adds lead time (typically 1–2 days per cycle, plus drilling and plating time), and each cycle introduces thermal stress on already-processed inner layers. Fabricators must qualify their press cycles for each material combination to ensure that the inner layer copper and laminate do not degrade during subsequent press cycles at elevated temperature and pressure.
In a 28-layer board assembled through three press cycles, cumulative layer-to-layer misregistration is a critical quality control challenge. Registration accuracy of ± 50 μm per layer, accumulated across 28 layers and three lamination cycles, demands:
High layer count and HDI complexity have a direct impact on yield and unit cost:
The manufacturing complexity of these boards is one reason that AI server PCB production is concentrated among a small number of tier-1 PCB fabricators with the capital equipment, process expertise, and quality systems required.
| Parameter | Standard Server Board (12L) | H100 HGX Baseboard (22L) | B200 Baseboard (28L) | NVSwitch 4.0 Board (36L) |
|---|---|---|---|---|
| Total layers | 12 | 22 | 28 | 36 |
| Signal routing layers | 6 | 8–10 | 10–12 | 14–18 |
| Power plane layers | 2 | 4–6 | 6–8 | 4–6 |
| Ground/reference layers | 4 | 8–10 | 10–12 | 14–16 |
| HDI build-up layers per side | 0 | 1–2 | 2–3 | 2–3 (or ELIC) |
| Lamination press cycles | 1 | 2–3 | 3–4 | 4–5 |
| Signal layer laminate | Standard FR4 | Megtron 6E / Tachyon 100G | Megtron 7 | Megtron 7 / HVLP copper |
| Minimum trace/space | 100 μm / 100 μm | 75 μm / 75 μm | 75 μm / 75 μm | 50–75 μm / 50–75 μm |
| Via-in-pad (VIPPO) | No | Yes | Yes | Yes |
| Backdrilling required | No | Yes (NVLink 4.0 + PCIe Gen5) | Yes (NVLink 5.0 + PCIe Gen6) | Yes / replaced by microvias |
| Typical fabrication lead time | 5–10 days | 15–20 days | 20–30 days | 30–40 days |
| Relative material cost | 1× | ~5–8× | ~10–15× | ~15–25× |
Is there a practical upper limit to how many layers an AI GPU PCB can have?
Yes, though it is set by manufacturing physics rather than an arbitrary rule. As layer count increases, board thickness increases, aspect ratio (board thickness / via diameter) rises, and it becomes progressively more difficult to achieve reliable copper plating in deep through-holes. At approximately 40+ layers with through-hole vias, aspect ratios exceed 20:1, and plating uniformity in the via barrel degrades below IPC Class 3 requirements. Any-layer HDI (ELIC) avoids this problem by eliminating full-board-thickness through-holes entirely, replacing them with stacked microvias. ELIC effectively removes the practical layer count ceiling imposed by through-hole plating constraints.
Do all GPU PCBs need 30+ layers, or just training boards?
Not all GPU PCBs need 30+ layers. Consumer GPU cards (RTX 4090) use 8–12 layers. Data center inference GPUs in PCIe add-in card format (L40S, A10G) use 12–16 layers. The 30+ layer count is specific to training-focused AI server baseboards that carry multiple GPU packages, NVSwitch chips, and route full NVLink 4.0 or 5.0 fabric connections. See AI Training vs AI Inference: Why They Need Different PCB Designs for a comparison of layer count requirements across deployment types.
What is the minimum layer count for an H100 SXM5 baseboard?
A functional H100 SXM5 baseboard with 8 GPUs and 4 NVSwitch 3.0 chips requires a minimum of approximately 18–20 layers to route all NVLink 4.0 connections, provide adequate power planes, and escape the BGA footprints of all packages. Designs at the low end of this range make aggressive routing compromises (shared routing channels, tighter spacing) that may compromise signal margins; 22–24 layers is the more practical minimum for a production-quality H100 baseboard that meets all signal integrity specifications with adequate margin.
Why can't AI GPU boards use fewer layers with denser routing per layer?
Two constraints prevent simply routing more traces per layer. First, the crosstalk specification between NVLink differential pairs limits how closely traces can be packed on a single layer; at 100–200 Gb/s per lane, near-end crosstalk (NEXT) must be < −30 dB, which requires a minimum edge-to-edge spacing of approximately 2× trace width between adjacent pairs. Increasing routing density beyond this spacing violates the crosstalk budget. Second, each signal layer must have an adjacent reference plane; eliminating reference planes to add more signal layers degrades impedance control and crosstalk simultaneously.
How does HDI reduce the required layer count compared to through-hole-only designs?
HDI does not necessarily reduce total layer count—it adds HDI build-up layers. However, it dramatically increases the routing density that can be achieved with a given number of layers, and it reduces the disruption to signal layers caused by through-hole via anti-pad clearances. Without HDI, routing from the inner rows of a fine-pitch GPU BGA would require so many through-hole vias that the anti-pad clearances in adjacent planes would effectively disrupt the reference plane beneath the critical NVLink routing—degrading impedance control across the entire BGA escape zone.
Which PCB fabricators can manufacture 30+ layer HDI boards for AI servers?
Production-capable 30+ layer HDI fabrication for AI server boards is concentrated among tier-1 fabricators with sequential lamination presses rated for low-loss specialty laminates (Megtron 7, Tachyon 100G), laser drill systems capable of 75 μm microvia diameter at the required stacking depth, controlled-depth backdrill CNC equipment, and fully automated layer registration verification. NextPCB is equipped to handle AI server board fabrication across this complexity range, from 16-layer H100 era boards through 32+ layer B200 and NVSwitch designs.
Whether you are fabricating a 20-layer H100 baseboard or pushing the limits of a 36-layer any-layer HDI NVSwitch board, NextPCB provides the sequential lamination capability, low-loss laminate processing, laser drilling, via-in-pad, backdrilling, and IPC Class 3 quality standards required for AI server PCB production.
Related Articles:
Still, need help? Contact Us: support@nextpcb.com
Need a PCB or PCBA quote? Quote now
