Infineon eFuses for AI Server Power Applications

­1. The birth of eFuses

AI servers are among the most demanding pieces of modern data-center equipment. Packs of GPUs, TPUs, or NPUs, ultrafast local NVMe storage, high-density memory, and massive switching fabrics push per-server power budgets into the kilowatt range. Those power rails are high-current, transient, and tightly sequenced, making them extremely sensitive; a single short, overload, or poor hot-swap can take down a multi-GPU board or, worse, damage an accelerator that costs more than many commodity servers.

These systems were historically protected by a mix of fuses, circuit breakers, discrete MOSFETs with controller ICs, current sense amplifiers, and dedicated hot-swap controllers. The approach works, but as server densities and complexities increase, so do the disadvantages – board area, parts count, calibration complexity, slower response times, inconsistent protection behavior, and difficult telemetry for power management.

Enter the eFuse – a family of integrated protection ICs that consolidate a hot-swap controller, a power MOSFET, and a current sensor into a single device. In the context of AI servers, where reliability, serviceability, and sophisticated power management matter as much as raw performance, eFuses are increasingly attractive. Infineon’s new XDP™ eFuses allow engineers to implement fast, accurate, and programmable electronic protection with lower BOM complexity, improved diagnostics, and a smaller thermal and PCB footprint.

2. eFuses in AI power applications

2.1 The power profile of AI servers

Modern AI workloads create extreme electrical demands such as high sustained currents, large inrush currents, fast transients, as well as strict sequencing and redundancy. These requirements stress both performance and protection elements, requiring that devices tolerate high steady currents without excessive loss while being capable of fast and accurate overcurrent/short protection.

2.2 Limitations of traditional protection

Traditional approaches — mechanical fuses, discrete MOSFETs with external controllers, or simple circuit breakers — have shortcomings for AI server use:

• Mechanical fuses are single-use and slow to clear, with no telemetry and a manual replacement cost and service outage.
• Discrete MOSFET + controller designs increase BOM and layout complexity. Multiple ICs, sense resistors, and gate drivers complicate routing and thermal design, and each interface is a potential failure mode.
• Current sense amplifiers plus comparators may provide detection but not safe, coordinated disconnect behavior – and they often require a separate gate-drive stage.
• Slow reaction times for large-scale mechanical devices or poorly tuned circuits can allow damaging energy to be delivered before the fault clears.

For AI servers that must be serviced quickly (often via hot-swap) with minimal downtime, these approaches are increasingly inadequate.

2.3 What an Infineon eFuse does

An Infineon eFuse integrates the three core elements needed for robust electronic protection (Figure 1):

• Hot-swap controller: Allows plugging a board or module into a live backplane without generating harmful transients; provides SOA (Safe Operating Area) controlled turn-on to limit inrush.
• Power MOSFET: Actual switching element that connects or disconnects the load under control.
• Current sensor: Precision sense amplifier or an integrated resistor that measures load current to detect overcurrent, short-circuits, or leaks.

Combined, the Infineon eFuse manages controlled soft-start during hot plugging, actively monitors current, trips rapidly on faults, provides thermal and overvoltage protections, and reports diagnostic and telemetry data over status pins or PMBus^®. This consolidated functionality replaces discrete hot-swap controllers and MOSFETs, bringing in compactness.

2.4 eFuse helps with hot-swap and serviceability in data centers

Hot-swapping of modules is critical to minimizing downtime but doing it haphazardly can cause more harm than good. But Infineon eFuses simplify that process by using a closed-loop controlled MOSFET gate, so that inrush current is limited to fully respect the MOSFET’s SOA, preventing large voltage sags on the backplane that would otherwise affect neighbor boards.

Similarly, in case of a fault, the eFuse can open quickly and isolate the faulty board without disrupting the rest of the chassis.And no matter the kind of fault, the eFuse provides solid status outputs such as open-drain outputs indicating overcurrent, over/undervoltage and thermal faults, enabling the management controller to rapidly identify and log failures for quick replacement.

2.5 Power distribution and efficiency

Since AI servers often use centralized DC (Direct Current) distribution (for example, 12 V or 48 V rails) feeding point-of-load converters, deploying eFuses at each module entry point gives logical, localized protection without complicated discrete topologies.

• Reduced parasitic losses and board area compared to using large discrete power MOSFETs and sense resistors scattered across the board.
• Better thermal behavior when the MOSFET is selected and matched to the application inside the package, often enabling simplified thermal management.

Table 1 compares the features of Infineon eFuses XDP730 and XDP72x.

Click image to enlarge

Table 1: Features comparison for Infineon XDP7x eFuses

3. Benefits of Infineon eFuses

One of the principal advantages is speed and determinism. Integrated current sensors and controllers allow eFuses to detect and clear faults in microseconds to a few milliseconds depending on their architecture. That speed matters when protectingexpensive accelerators that can be damaged by short, high-energy faults, as well as downstream circuitry, such as VRMs and high-value memory DIMMs.

Faster clearing reduces energy dissipated during the fault and limits collateral damage.

Moreover, by integrating controllers, MOSFETs, and sensing, eFuses reduce BOM complexity. Fewer discrete components imply fewer drivers, sense amps, and passives. This simplifies routing by removing the need for separate sense traces or Kelvin sense routing, assuming the eFuse provides integrated or carefully defined sense points. Majorly, it creates a smaller footprint enabling tight server boards where every mm² counts.

Fewer parts also mean fewer suppliers to qualify, fewer assembly steps, and decreased failure modes.

In addition to advantages in reactive maintenance, Infineon eFuses provide programmable behavior:

• Adjustable trip thresholds (current, voltage, and thermal).
• Selectable trip response (fast shutdown, auto-retry, and latch-off).
• Configurable inrush control (fully digital SOA control).

This programmability enables the same eFuse family to be used across multiple rails or platforms, saving development time and simplifying spares management. It also enables runtime adaptability since power management firmware can temporarily relax limits for peak loads or set tighter limits during known maintenance windows.

Such proactive maintenance goes one step further with detailed telemetry which is a big deal in modern data centers:

• Operational metrics: Digital readouts of current, voltage, power, and energy, which can feed into data-center management systems for predictive maintenance. The current can also be monitored by one analog pin IMON.
• Event logging: Report fault type and peak/valley current for trip history via status pins or PMBus^®.
• Quicker troubleshooting: Clearer diagnostics reduce mean time to repair (MTTR) for field technicians.

When it comes to costs as well, even though eFuses cost more per unit than a simple fuse, they reduce lifecycle costs, improving TCO as manual fuse replacements are eliminated to reduce downtime. They lower the risk of collateral damage to expensive modules and reduce spare mechanical fuses in stockrooms while also reducing technician visits.

Because eFuses combine MOSFET and package optimizations, they can:

• Provide low R_DS(on) MOSFET matched to the package thermal characteristics, reducing conduction losses.
• Use package thermal pathways to better dissipate heat vs. a distributed discrete MOSFET plus sense resistor solution.
• Improve thermal predictability, which simplifies chassis cooling design for tightly packed servers.

This matters greatly when multiple high-current rails are concentrated on the same board.

Click image to enlarge

Figure 2: EVAL_XDP730-001

Click image to enlarge

Figure 3: EVAL_XDP72X-001

4. Practical application examples of Infineon eFuses

To make this concrete, here are a few ways Infineon eFuses can be used in AI server designs:

1. GPU module protection: Place eFuses on each GPU power input to isolate a failed GPU without taking down the entire board. The eFuse provides inrush control during hot plugging and fast disconnect on short.
2. NVMe sleds/hot-swap bays: eFuses on each drive bay enable individual drive isolation and provide current telemetry for predictive failure detection.
3. VRM protection: eFuses upstream of VRMs can prevent catastrophic VRM failures from cascading into major system outages, and can provide current limiting to protect expensive capacitors and MOSFET stacks.
4. Redundant hot-swap PSUs: Using eFuses to measure and control current sharing between redundant supplies isolates a failing supply without impacting others.
5. Fan control and protection: Use eFuses to turn server fans on or off, measure current for feedback, and to report and protect if any fault is detected.

Each use case benefits from the eFuse’s ability to combine protective, sensing, and digital SOA control in startup in one compact device. Infineon also have evaluation boards for eFuses XDP730 and XDP72x. Figure 2 shows EVAL_XDP730-001 with maximum three XDP730 in parallel for higher current. Figure 3 shows EVAL_XDP72X-001 with one XDP720 and one XDP721 in parallel. For more information on Infineon eFuses, visit our website.