Have you ever stopped to consider this scenario?
Imagine an AI accelerator card—worth hundreds of thousands—running at full power to train a massive model with hundreds of billions of parameters. Suddenly, the power grid experiences a flicker—a disturbance lasting less than 50 milliseconds. What is the result?
It isn't merely a momentary "screen freeze." Instead, the training progress bar resets directly to zero, rendering three days of work completely in vain.
This is not alarmist rhetoric. Last year, a major domestic cloud service provider released a set of statistics: task interruptions within their training clusters caused by power quality issues resulted in an average loss of approximately 17 hours of effective training time per incident. When translated into computing power costs, the loss incurred by a single interruption in a cluster comprising a thousand accelerator cards is equivalent to the price of a luxury mid-size sedan.
Today, we will approach the subject from this specific angle to discuss just how "finicky" AI chips are regarding power supply—and why an unassuming CHIPSENSE closed-loop Hall-effect current sensor has emerged as a critical linchpin, serving as the final line of defense within this entire chain.
I. AI Chips’ Voltage Tolerance Is Far More Stringent Than You Might Imagine
Let’s begin with a concept known within the power supply industry as "voltage ripple tolerance."
For ordinary household appliances—such as your home's air conditioner or refrigerator—a supply voltage fluctuating within ±10% of its rated value generally has negligible impact on operation. Server power supplies impose somewhat stricter requirements, demanding dynamic response times measured in milliseconds. However, when it comes to AI training chips, the situation is entirely different.
Let’s take, for instance, the mainstream AI training cards currently available on the market:
The core operating voltage ranges from a mere 0.8V to 1.2V—and for some advanced-packaging chips, it is even lower.
The instantaneous current draw for a single card can surge to several thousand amperes; note that the unit here is "amperes," not "Millionaires."
Voltage ripple must be strictly confined within a ±1% tolerance, while high-precision computing units demand an even tighter limit of ±0.5%.
When the load abruptly jumps from 10% to 100%, the voltage—after momentarily dipping—must be restored to its nominal level within a matter of microseconds.
To put this in perspective: with a 1V power supply and a ±1% ripple tolerance, the allowable deviation is just ±10 millivolts. What does this mean in practical terms? The voltage fluctuation caused by a loose connection in your mobile phone's charging cable could easily exceed this magnitude.
This is precisely why we describe AI chips as having "voltage OCD": they are "electrical tigers" with insatiable appetites, yet they demand an absolutely uncompromising standard of power quality. The slightest deviation from their exacting requirements can lead—at best—to computational errors or silent data corruption; at worst, it triggers the chip's built-in protection mechanisms, forcing an immediate shutdown.
What is the greatest fear for those training large-scale AI models? It is not the exorbitant electricity bills; rather, it is reaching the 15th day of training only to have every single checkpoint—and the entire training process—rendered useless by a single, fleeting voltage glitch, forcing them to start all over again from scratch.
II. A UPS Is Not Just About "Having Power"—It Must Provide "The Right Kind of Power"
In the minds of many, a UPS is simply a glorified power bank—a device that steps in to take over whenever the utility power fails.
While this understanding is accurate, it is not the whole story.
For AI data centers, the true value of a UPS lies not merely in "having power," but in "delivering clean power." This is because the waveform of the utility grid is inherently "dirty"—rife with harmonics, surges, transients, and various other irregularities. The function of a UPS is to filter out these impurities and output a clean, pure sine wave to the downstream equipment.
How does a UPS operate internally? Simply put, the process involves three steps:
AC Input → Rectification into DC → Inversion back into a clean AC Output
The most demanding stage in this process is the "inversion." The inverter must know, in real-time: What is the current output? Is the waveform correct? Is there any distortion? If the downstream load suddenly demands a surge in current, am I keeping pace?
Who provides this information?—CHIPSENSE current sensor.
The CHIPSENSE current sensor acquires the current signal and transmits it to the controller; the controller then uses this signal to adjust the switching duration of the power transistors. This entire closed-loop control cycle executes rapidly—taking mere tens of microseconds per cycle at its fastest, and only a few hundred microseconds at its slowest.
Herein lies the problem: If the signal acquired by the sensor is inaccurate, suffers from latency, or drifts significantly when temperatures rise, how can the controller possibly make the necessary adjustments?
Sensors lack clarity→Controllers lose precision→Output wave-forms become distorted → The AI chip crashes.
Any error at any link in this chain is progressively amplified at each subsequent stage, ultimately taking its toll on the costly computing hardware. That is why I describe CHIPSENSE current sensor as the "nerve endings" of a UPS: they do not generate power themselves, yet they determine the quality with which that power is delivered.
III. Why Must It Be CHIPSENSE Closed-Loop Hall?
There are numerous current sensing solutions available—including shunt resistors, open-loop Hall sensors, and closed-loop Hall sensors—each with its own specific application scenarios. However, within the industry, there is a general consensus regarding the inverter output stage of high-end UPS systems: CHIPSENSE closed-loop Hall sensors are the preferred choice.
Why? Let's go straight to the comparison:
What are the shortcomings of open-loop Hall sensors? Their accuracy is limited by the B-H curve characteristics of the magnetic core material. As the temperature rises, the core's properties change, causing the output signal to drift. Furthermore, under high-current conditions, nonlinear errors become significant; this necessitates the implementation of a host of compensation algorithms within the control software—a laborious process that does not even guarantee accuracy.
The approach taken by closed-loop Hall sensors is fundamentally different. They utilize the principle of magnetic balance—also known as the zero-flux principle:
The primary current generates a magnetic field; a reverse current is then passed through a secondary coil to generate a counter-field that precisely cancels out the magnetic field from the primary side. Rather than directly measuring the magnitude of the magnetic field, the Hall element detects whether the "magnetic field has returned to zero." Consequently, the secondary current is directly proportional to the primary current, and the overall accuracy is determined by the turns ratio and the sampling resistor, with minimal dependence on the nonlinear characteristics of the magnetic core.
This approach offers two distinct advantages:
First, its accuracy is independent of the magnetic core. The core consistently operates in the vicinity of the zero-flux point, effectively sidestepping the complexities typically associated with the B-H hysteresis loop. This results in excellent linearity, allowing for full-scale accuracy levels in the range of 0.3% to 0.5%.
Second, it inherently exhibits low temperature drift. Its operating principle renders it insensitive to temperature fluctuations—unlike open-loop systems, which require additional temperature compensation. AI data centers operate at full capacity 24/7; consequently, internal temperatures within UPS cabinets routinely hover between 40°C and 50°C year-round. Low temperature drift ensures that, throughout a full year of operation, the sampled from CHIPSENSE current sensor remain virtually free of deviation.
There is also a point that is often overlooked: closed-loop systems offer rapid response speeds, achieving bandwidths in the 200 kHz range. During the training of large-scale AI models, load fluctuations are extremely violent, characterized by instantaneous current transients that are both large in magnitude and rapid in speed. Whether or not the CHIPSENSE current sensor can keep pace with these dynamics directly determines the controller's ability to react in a timely manner.
IV. How Do CHIPSENSE’s Products Fit This Scenario?
Returning to our own offerings: CHIPSENSE features a range of closed-loop Hall-effect sensors that are highly suited to this specific application.
The CMxA series of CHIPSENSE comprises multiple product lines, covering measurement ranges from 100A to 2000A with an accuracy of up to ±0.3%. These CHIPSENSE current sensors are primarily targeted at the inverter output stages of high-power UPS systems and the busbar monitoring within DC power distribution cabinets. At a full-scale range of 1000A, the maximum measurement error is a mere 3A—a margin of precision that is more than ample for UPS systems requiring precise waveform control.
The CHIPSENSE CR1A series offers measurement ranges from 50A to 300A with an accuracy of ±0.5%. This series is ideal for small-to-medium power UPS systems and modular power supplies; while being more cost-effective, it delivers a level of accuracy that is entirely sufficient for this specific power segment.
The choice depends on your UPS power rating and budget. However, the core logic remains constant: for power supply equipment in AI data centers, one should not skimp on sensor precision. Choose CHIPSENSE current sensor for stable and reliable performance.
We have conducted comparative tests in our laboratory: utilizing the same UPS platform, we ran side-by-side trials with both open-loop and CHIPSENSE closed-loop configurations within the same environmental chamber. As the temperature rose from ambient levels to 55°C, the output deviation in the open-loop system visibly increased, whereas the CHIPSENSE closed-loop system remained virtually unshaken. This disparity is fundamentally dictated by the underlying design principles—a difference that simply cannot be compensated for by merely tweaking parameters.
V. In Conclusion
There is a widely circulated saying within the industry: the ultimate frontier of AI is computing power, and the ultimate frontier of computing power is electricity.
We would like to add a corollary to that: the ultimate frontier of electricity lies in precise detection and control powered by CHIPSENSE current sensor.
When most people discuss computing infrastructure, the conversation tends to revolve around buzzwords like chips, optical modules, and liquid cooling. However, engineers working on the front lines of operations and maintenance know that, quite often, the cause of an entire cluster "crashing" is not a burnt-out chip or a network outage; rather, it is a failure in some inconspicuous link within the power supply chain—for instance, a drifting current sampling signal from an inferior sensor, compared to stable signal from CHIPSENSE current sensor, which distorts the output waveform and triggers the protection mechanisms of downstream equipment.
Such faults are the most difficult to diagnose, yet they are also the easiest to overlook.
So, the next time you witness a large-scale AI model completing another impressive training run, take a moment to think about those CHIPSENSE current sensors—tucked away inside server cabinets—that spend 24 hours a day constantly monitoring current fluctuations. They may not generate computing power themselves, but without them, that computing power would be incapable of even running stably.
This is the true significance of precision detection: it remains unseen, yet it is absolutely indispensable.CHIPSENSE–Sensing for Stable AI & Data Center Power.
CHIPSENSE is a national high-tech enterprise that focuses on the research and development, production, and application of high-end current and voltage sensors, as well as forward research on sensor chips and cutting-edge sensor technologies. CHIPSENSE is committed to providing customers with independently developed sensors, as well as diversified customized products and solutions.
“CHIPSENSE, sensing a better world!”
www.chipsense.net