12.2a PCIe Edge Cards: standard slot insertion, passive cooling, TDP ≤350W¶
🌐 Context Introduction¶
When you're starting out with AI infrastructure, one of the first physical components you'll encounter is the PCIe Edge Card. This is the most common form factor for NVIDIA GPUs used in data centers. Unlike the bulky graphics cards in gaming PCs, these cards are designed to slide directly into standard server slots, require no fans (passive cooling), and stay within a power envelope of 350W or less. Understanding this form factor is essential because it dictates how you rack, power, and cool your AI servers.
⚙️ What is a PCIe Edge Card?¶
A PCIe Edge Card is a GPU that connects to a server's motherboard via the PCI Express (PCIe) slot. Think of it like a high-performance expansion card — similar to a network card or storage controller, but much more powerful.
Key characteristics: - Standard slot insertion: Fits into any standard PCIe slot (typically x16) on a server motherboard. - Passive cooling: No fans attached. Heat is removed by the server's chassis airflow (system fans). - TDP ≤ 350W: Thermal Design Power is capped at 350 watts, ensuring compatibility with standard server power and cooling designs.
🛠️ Physical Design & Installation¶
📐 Form Factor Dimensions¶
- Full-height, full-length (FHFL): Most common size for data center PCIe GPUs.
- Dual-slot width: Occupies two adjacent PCIe slots to accommodate the heatsink.
- Standard bracket: Includes a metal bracket that screws into the server chassis.
🔌 Insertion Process (Simplified)¶
- Power off the server and ground yourself to avoid static discharge.
- Align the GPU with the PCIe slot on the motherboard.
- Press firmly until the retention clip clicks into place.
- Secure the bracket with screws to the chassis.
- Connect power cables (typically one or two 8-pin PCIe power connectors).
- Close the server and power it on.
🧠 Tip: Always check the server's PCIe slot power rating. Some slots provide only 75W, so the GPU will need additional power from the PSU.
❄️ Passive Cooling Explained¶
Unlike consumer GPUs with visible fans, data center PCIe Edge Cards rely on passive cooling. This means: - No onboard fans — reduces noise and moving parts. - Large heatsink — a finned metal block that absorbs heat from the GPU chip. - Server airflow — the server's front-to-back fan array pushes air over the heatsink.
💨 Airflow Requirements¶
- Front-to-back flow: Air enters from the server front, passes over the GPU, and exits at the back.
- Minimum CFM: Each GPU model specifies a minimum cubic feet per minute (CFM) of airflow.
- Temperature limits: GPU junction temperature must stay below ~85°C under load.
⚠️ Warning: Never operate a passive-cooled GPU outside a server chassis with forced airflow. It will overheat in seconds.
📊 Visual Representation: PCIe Edge Card GPU Installation¶
This diagram displays standard PCIe edge card installation on server motherboards, communicating via the PCIe slot bus.
🔋 Power Considerations (TDP ≤ 350W)¶
TDP (Thermal Design Power) is the maximum heat the cooling system must dissipate. For these cards, it's 350 watts or less.
📊 Power Delivery Comparison¶
| Feature | Standard PCIe Slot | PCIe Edge Card (≤350W) |
|---|---|---|
| Slot power (max) | 75W | 75W (from slot) |
| Additional power needed | N/A | 225–275W via cables |
| Power connectors | None | 1x 8-pin (150W) or 2x 8-pin (300W) |
| Typical PSU requirement | 500W | 800W–1000W per server |
🔌 Power Cable Configuration¶
- Single 8-pin: Provides up to 150W (total 225W with slot).
- Dual 8-pin: Provides up to 300W (total 375W with slot — but limited to 350W by design).
- 12VHPWR: Newer cards may use a single 12+4 pin connector (up to 600W, but limited to 350W here).
🕵️ Common NVIDIA PCIe Edge Cards (≤350W TDP)¶
| GPU Model | TDP | Cooling | Slot Width | Use Case |
|---|---|---|---|---|
| NVIDIA A2 | 60W | Passive | Single-slot | Edge inference |
| NVIDIA L4 | 72W | Passive | Single-slot | AI video & inference |
| NVIDIA A16 | 250W | Passive | Dual-slot | Virtual desktop (VDI) |
| NVIDIA A30 | 165W | Passive | Dual-slot | Enterprise AI |
| NVIDIA L40S | 350W | Passive | Dual-slot | AI training & inference |
✅ Key Takeaways for New Engineers¶
- PCIe Edge Cards are the standard GPU form factor for most AI servers.
- Passive cooling means the server's system fans do all the work — never run these cards outside a chassis.
- TDP ≤ 350W ensures compatibility with standard server power supplies and cooling designs.
- Always verify the server's PCIe slot generation (Gen4 vs Gen5) and power delivery before installation.
- Airflow direction matters — install the GPU so its heatsink fins align with the server's airflow path.
📚 Further Learning¶
- Review your server's technical manual for PCIe slot power limits and airflow specifications.
- Check the NVIDIA datasheet for your specific GPU model to confirm TDP and cooling requirements.
- Practice ESD-safe handling — GPUs are sensitive to static electricity.
💡 Remember: A properly installed PCIe Edge Card should be silent (no fan noise), cool (within thermal limits), and draw power cleanly from the server's PSU. If you hear fans or see thermal throttling, something is wrong with the airflow or power delivery.