10%-20% of computing power is being consumed by electromagnetic pollution—who foots the bill?
简体中文
7days a week from 9:00 am to 9:0pm
+86 18016225001

Shanghai Leiditech上海雷卯电子科技有限公司是专业的静电保护元件厂家,TVS二极管供应商;专业提供防雷防静电方案,电磁兼容EMC免费测试等服务,品质保证,库存充足,型号齐全,值得信赖,如有采购静电保护元件,TVS二极管需求,请联系雷卯,24小时服务热线:021-50828806.

By LEIDITECH | 31 August 2026 | 0 Comments

10%-20% of computing power is being consumed by electromagnetic pollution—who foots the bill?

 

 

10%-20% of computing power is being consumed by electromagnetic pollution—who foots the bill?

Let’s do the math first.

A large-scale model training project involves tens of millions in GPU cluster procurement, millions in monthly electricity costs, and a team of over a dozen for operations. Then you realize the actual computing power is only 80%-90% of the nominal value.

Where did the missing 10%-20% go?

It’s not a GPU quality issue, not insufficient power supply, not network bandwidth bottlenecks — it’s electromagnetic interference, a hidden cost item that doesn’t even appear on financial statements.

 

10%-20% computing power loss = millions more wasted each year

 

Let’s break this down further.

Assume an AI training cluster purchases 100 H100 GPUs, with a single-card cost of approximately 250,000 RMB, total investment of 25 million RMB. If electromagnetic interference causes a 15% computing power loss, that’s equivalent to 15 H100 GPUs — worth 3.75 million RMB — running uselessly.

Add to that the issues caused by EMI:

· Training interruption and checkpoint restart: Each restart loses at least several hours of training time; for a large model, a single training cycle can cost millions.

· Accelerated hardware aging: Frequent signal anomalies accelerate wear on chips and power modules, shortening equipment lifespan by 20%-30%.

· Increased operational overhead: Troubleshooting "ghost faults" consumes over 30% of O&M team working hours.

· Secondary incidents caused by sensor failures: Thermal throttling, equipment burnout, emergency termination of training tasks.

Adding these up, the hidden annual loss due to EMI in a medium-sized AI data center is conservatively estimated at over 10 million RMB.

And the investment required to solve this? The component cost for an EMC protection solution may account for less than 1% of the GPU procurement budget.

Who should pay for this "hidden tax"?

Hardware vendors? AI accelerators are designed with performance and power consumption as top priorities. EMC protection is often a "patch job" at the tail end of PCB design. Chip frequencies have broken through the GHz barrier, high-frequency switching power supplies switch millions of times per second, each time radiating broadband electromagnetic signals — but nowhere on the hardware datasheet does it state "this product may interfere with your wireless sensors."

Data center builders? Traditional data center shielding standards were based on general-purpose computing equipment from a decade ago. AI rack electromagnetic field intensity has long exceeded the original thresholds. Chassis ventilation holes and cable openings have become antennas for electromagnetic signals. But builders construct to old standards and pass inspections by old standards.

Operations teams? They can see temperature curves, power consumption data, and network latency — but they cannot see electromagnetic spectrum plots. Conventional O&M tools cannot detect electromagnetic interference; "ghost faults" can only be resolved by rebooting.

The root of the problem: Three parties are siloed, and no one takes responsibility. Hardware designers do not understand the data center environment; data center builders do not understand electromagnetic compatibility; operations teams lack EMC detection tools. When computing power inexplicably drops by 10%-20%, no one can clearly explain "why," and no one knows "who to contact."

 

️Leiditech: Turning EMC Protection into "Insurance", Not "Compensation"

 

Leiditech's 15 years of EMC practical experience have distilled one iron rule: the cost of rectifying electromagnetic interference is 8-10 times the cost of front-end protection. The earlier you intervene, the lower the cost.

 

 

Layer 1: Power Protection — Install "Safety Valves" on Power Rails

AI server power rails face dual threats of transient surges and switching power supply noise. Leiditech's high-reliability TVS diodes (6600W, -55°C to 175°C wide temperature, batch parameter deviation ≤±5%) with low leakage current and low clamping voltage, combined with PPTC resettable fuses, form a "clamping + overcurrent" dual protection scheme. Verified in 7×24 uninterrupted operation, leakage current is less than 1µA, with zero power consumption loss under long-term loading. The accompanying common-mode choke (LDW43T-513T) suppresses power common-mode noise, validated in over 50 projects.

Layer 2: Signal Protection — Install "Filters" on High-Speed Buses

Leiditech's low-junction-capacitance ESD series (starting from 0.05pF) ensures signal attenuation ≤0.5%. TLP measured data: under 16A surge, clamping voltage is 9.4V, dynamic resistance 0.3Ω, outperforming international counterparts. For advanced-process AI chips with a withstand voltage of 10V, this is a non-negotiable survival baseline.

Layer 3: PCB Layout — Eliminate Parasitic Risks from the Source

"3mm Rule": Place protection devices within 3mm of the connector, reducing parasitic inductance by 70%. "Symmetric Routing": Trace length matching to data lines and ground within ≤0.5mm, ensuring surge response <1ns. These are not theoretical derivations, but engineering conclusions drawn from thousands of measurements in Leiditech's laboratory.

Layer 4: Verification Closed Loop — In-House EMC Laboratory

 

ESD 30kV, EFT 4kV, surge (8/20, 10/700, 10/1000) — full test suite completed in one stop. Solution design is verified immediately upon completion, eliminating the wait for third-party queues, compressing the rectification cycle from weeks to days.

 

In the era of computing power, EMC protection is the investment with the highest ROI.

 

When a single H100 costs 250,000 RMB and a single training cycle costs millions, spending a few tens of thousands on EMC protection may yield the highest ROI of any investment in the entire data center.

The question is not "whether to do it," but "who realizes the need to do it first." Those teams still treating 10%-20% computing power loss as "normal attrition" are effectively spending millions more each year on a problem that could have been solved.

The IEC 62368-3:2026 standard has been released; Apple, NVIDIA, and Huawei Ascend have already followed suit. Domestic AI data centers should also wake up — every fraction of lost computing power is real money.

Core argument: The 10%-20% computing power loss in AI data centers is not "normal fluctuation" but a hidden cost caused by electromagnetic interference, with annual losses reaching millions. The root cause lies in the fragmentation among hardware design, data center construction, and operations management, with no one taking responsibility. Leiditech's full-link EMC protection solution provides a systematic fix at an extremely low investment ratio (<1% of computing power budget).

#AIcomputingpowerloss #ElectromagneticInterference #Leiditech #EMCprotection #DatacenterCost #TVSdiode #ComputingROI #ElectromagneticCompatibility #Localization #GPUcluster

Leave a Reply

Your email address will not be published.Required fields are marked. *
Name
E-mail
Content
Verification code
ver_code