11.2b GDDR physical packaging: mounted chips on PCB, width limited by board lanes

📦 Nvidia GPU Architecture 📖 GPU Memory Subsystems

🧠 Context Introduction

When you look at a modern GPU, you'll notice several small black chips arranged around the main processor. Those are GDDR memory chips — the high-speed VRAM that stores textures, frame buffers, and AI model weights. But how these chips are physically attached to the circuit board (PCB) matters a lot for performance. This topic explains the physical packaging of GDDR memory: how chips are mounted on the PCB and why the memory bus width is limited by the number of board lanes available.


⚙️ How GDDR Chips Are Mounted on the PCB

GDDR memory chips are surface-mounted directly onto the printed circuit board (PCB). Each chip is a small, rectangular package with dozens of tiny solder balls underneath (called a BGA — Ball Grid Array). These solder balls connect the chip's internal circuits to the copper traces on the PCB.

Key points about mounting: - Each GDDR chip has its own dedicated data bus (typically 32 bits wide per chip). - Multiple chips are placed around the GPU die, usually in a symmetrical pattern. - Traces (thin copper wires) run from each memory chip to the GPU's memory controller on the die. - Signal integrity is critical — longer traces or poorly routed traces can cause data errors at high speeds.


📊 The "Width" Problem: Why Board Lanes Are the Bottleneck

The memory bus width (e.g., 256-bit, 384-bit) is the total number of bits the GPU can read from or write to memory in a single clock cycle. This width is determined by:

  • Number of GDDR chips × bits per chip = total bus width

For example: - 8 chips × 32 bits each = 256-bit bus - 12 chips × 32 bits each = 384-bit bus

But here's the catch: the PCB has a limited number of physical lanes (copper traces) that can run from the GPU to the memory chips without causing interference or signal degradation. This limitation comes from:

  • PCB layer count — more layers cost more money and increase complexity.
  • Trace length — longer traces introduce latency and signal loss.
  • Crosstalk — adjacent traces can interfere with each other at high frequencies.
  • Power delivery — more memory chips require more power and more decoupling capacitors.

🛠️ Physical Constraints in Practice

Constraint Impact on Memory Width
PCB size Larger boards can fit more chips, but cost and form factor limit this.
Number of PCB layers More layers allow more traces, but increase manufacturing cost.
Signal integrity High-speed GDDR (like GDDR6X) requires careful trace routing, limiting how many lanes can be packed together.
Thermal management More chips generate more heat, requiring better cooling solutions.
GPU die size The GPU's memory controller has a fixed number of physical pins — this sets the maximum number of lanes.

📊 Visual Representation: GDDR Physical Board Layout

This diagram shows the physical layout of GDDR memory chips placed around the central GPU die, connected via standard PCB traces.

flowchart LR GDDR1["GDDR Chip"] --- Trace1["PCB Traces"] GDDR2["GDDR Chip"] --- Trace2["PCB Traces"] Trace1 --> GPUDie["Central GPU Die"] Trace2 --> GPUDie classDef cpu fill:#eafaf1,stroke:#76b900,stroke-width:2px,rx:6px,ry:6px; classDef memory fill:#f0f7ff,stroke:#3498db,stroke-width:1.5px,rx:4px,ry:4px; classDef system fill:#f1f5f9,stroke:#64748b,stroke-width:1.5px; class GPUDie cpu; class GDDR1,GDDR2 memory; class Trace1,Trace2 system;

🕵️ Real-World Example: Comparing Memory Configurations

Let's look at two common GPU memory configurations to see how physical packaging affects width:

Entry-level GPU (e.g., RTX 4060): - 6 GDDR6 chips - Each chip: 32-bit bus - Total bus width: 6 × 32 = 192-bit - Chips are mounted on one side of the PCB, in a compact layout.

High-end GPU (e.g., RTX 4090): - 12 GDDR6X chips - Each chip: 32-bit bus - Total bus width: 12 × 32 = 384-bit - Chips are mounted on both sides of the PCB (some on front, some on back) to fit within the board's physical dimensions.


🔍 Why Can't We Just Add More Chips?

You might wonder: why not just put 16 chips for a 512-bit bus? Here's why that's difficult:

  1. PCB real estate — The GPU die is only so big; there's limited space around it for memory chips.
  2. Trace routing complexity — Each additional chip requires 32+ traces to the GPU. At high speeds, these traces must be exactly the same length to avoid timing skew.
  3. Power delivery — More chips draw more current, requiring thicker power planes and more capacitors.
  4. Cost — High-layer-count PCBs with complex routing are expensive to manufacture.

📈 Summary Table: GDDR Physical Packaging Trade-offs

Factor Benefit of More Chips Limitation
Bus width Higher bandwidth (more bits per cycle) PCB trace count and routing complexity
Capacity More VRAM for large models Physical space on PCB
Cost N/A More chips + complex PCB = higher cost
Power N/A More chips = higher power draw and heat

🧪 Key Takeaway for New Engineers

When you see a GPU spec sheet listing a 256-bit or 384-bit memory bus, remember that number isn't arbitrary — it's a direct result of how many GDDR chips the engineers could physically fit around the GPU die and reliably connect using the available PCB lanes. The memory bus width is a fundamental design constraint that balances performance, cost, power, and physical space.

As AI workloads demand more memory bandwidth, engineers are pushing the limits of GDDR packaging — using faster chips (GDDR6X), more layers in the PCB, and even stacking chips (3D packaging) to overcome the lane limitation.