2.1d UPS (Uninterruptible Power Supplies): online double-conversion vs line-interactive¶
🌐 Context Introduction¶
When you're working with AI infrastructure, power reliability is not optional—it's critical. A single power glitch can corrupt training data, crash a GPU cluster, or damage sensitive electronics. That's where Uninterruptible Power Supplies (UPS) come in. They act as a bridge between the utility grid and your equipment, providing clean, stable power during outages, sags, surges, and frequency fluctuations.
For new engineers, understanding the two most common UPS topologies—online double-conversion and line-interactive—is essential. They differ in how they handle power conditioning, efficiency, and battery usage. This guide breaks down the differences in simple terms.
⚙️ What is a UPS?¶
A UPS is a battery-backed power system that: - Provides immediate power when the main supply fails - Conditions the incoming power to remove noise, spikes, and dips - Gives you time to safely shut down equipment or switch to a generator
🛠️ Line-Interactive UPS¶
How it works: - Normally, the UPS passes utility power directly to your equipment. - It uses a voltage regulator (automatic voltage regulation, or AVR) to correct minor sags or surges without switching to battery. - When the power fails completely, it switches to battery power in 2–4 milliseconds.
Key characteristics: - ✅ More efficient (97–99% efficiency) because power flows directly through - ✅ Lower cost and smaller footprint - ❌ No isolation from utility power noise or frequency issues - ❌ Battery is used only during outages or severe fluctuations
Best for: - Office environments, workstations, network switches - Equipment that can tolerate brief transfer times - Budget-conscious deployments
⚡ Online Double-Conversion UPS¶
How it works: - Incoming AC power is converted to DC, then converted back to AC (hence "double-conversion"). - The battery is always connected to the DC bus, so there is zero transfer time to battery. - The output is completely regenerated—clean, stable, and isolated from the grid.
Key characteristics: - ✅ Perfect power conditioning—removes all noise, harmonics, and fluctuations - ✅ Zero transfer time—critical for sensitive AI hardware (GPUs, storage arrays) - ✅ Protects against frequency variations and waveform distortion - ❌ Lower efficiency (90–95%) due to constant conversion - ❌ Higher cost and generates more heat
Best for: - AI training clusters, GPU servers, HPC nodes - Data centers with sensitive or mission-critical loads - Environments with unstable grid power
📊 Comparison Table: Online Double-Conversion vs Line-Interactive¶
| Feature | Line-Interactive | Online Double-Conversion |
|---|---|---|
| Power conditioning | Basic (voltage regulation only) | Full (noise, spikes, frequency) |
| Transfer time to battery | 2–4 ms | Zero (always on battery) |
| Efficiency | 97–99% | 90–95% |
| Cost | Lower | Higher |
| Heat generation | Low | Moderate |
| Isolation from grid | No | Yes |
| Best use case | Workstations, switches | GPU clusters, AI servers |
🕵️ When to Use Which in AI Infrastructure¶
For AI workloads, here's a simple rule of thumb:
- Line-interactive is acceptable for:
- Management servers, monitoring stations, network gear
- Non-critical storage or backup nodes
-
Development workstations
-
Online double-conversion is strongly recommended for:
- GPU compute nodes (e.g., NVIDIA DGX systems, A100/H100 clusters)
- High-speed storage arrays (NVMe, all-flash)
- Any equipment running long training jobs (hours or days)
- Systems where a power glitch could corrupt data in memory
Why GPUs need online double-conversion: - GPUs draw high, variable power—line-interactive units may struggle with voltage regulation - A 2 ms transfer gap can cause GPU memory errors or training job failures - Double-conversion provides clean sine wave output, which modern PSUs prefer
📊 Visual Representation: UPS Topology Comparison¶
This diagram illustrates the difference between Line-Interactive and Online Double-Conversion UPS systems, highlighting why sensitive GPU compute loads require the zero-transfer-time, continuous regeneration pathway.
🔌 Real-World Example¶
Imagine you're setting up a small AI lab with: - 4x NVIDIA A100 GPUs in a server - 1x network switch - 1x management workstation
Recommended approach: - Put the GPU server on an online double-conversion UPS (e.g., 10–15 kVA) - Put the switch and workstation on a line-interactive UPS (e.g., 1500 VA) - This balances cost while protecting the most critical load
✅ Key Takeaways for New Engineers¶
- Line-interactive = efficient, affordable, good for less sensitive gear
- Online double-conversion = maximum protection, essential for AI compute
- Always check the load type—GPUs and storage arrays need clean, zero-transfer power
- Efficiency matters—online units waste more heat, so plan cooling accordingly
- Battery runtime is separate from topology—both can support extended runtimes with external battery packs
📚 Further Reading (for your own study)¶
- UPS sizing calculators (e.g., APC, Eaton, Vertiv)
- Power quality standards (IEEE 519, ITIC curve)
- NVIDIA DGX system power requirements documentation
Remember: In AI infrastructure, power is the foundation. A well-chosen UPS protects not just hardware, but the hours of compute time and data integrity behind every training run.