Recommended Free Tools
An AI server GPU throttles from a power limit when its power-management system lowers clock speeds to keep consumption inside an enforced electrical budget. Confirm it by correlating reduced clocks and throughput with a power-cap event and power draw near the effective ceiling. Then rule out thermal protection, an external power brake, application-clock settings, synchronized boost behavior, workload starvation, and server-level policy. A cap event alone does not mean the GPU is faulty, and raising a limit is not automatically safe.
What power-limit throttling actually means
Modern accelerator boards continuously balance performance against a configured power envelope. Under a sufficiently demanding workload, the GPU can reduce its SM (streaming multiprocessor) and memory clocks so instantaneous or sustained draw stays within that envelope. Throughput falls because fewer arithmetic operations are completed per second, memory bandwidth may decrease, or both.
As an Amazon Associate I earn from qualifying purchases.
NVIDIA describes the mechanism this way: “When power management is enabled the GPU limits power draw under load to fit within a predefined power envelope by manipulating the current performance state.” That is deliberate control behavior, not proof of a failed component. The practical question is whether the limit is the intended one for your platform and workload, or an unexpectedly restrictive policy, thermal condition, or power-delivery signal.
Power throttling must be diagnosed as a time-aligned relationship, not from one number. A low power reading while a job is waiting on data, blocked on synchronization, or nearly idle is normal. Look for a period in which utilization and workload demand are high, clocks fall, throughput or latency worsens, and a limit reason is asserted at the same time.
#1 Best Overall
- Ventilation Fan: Designed to quietly ASUS GT/RT- AC5300 , cool Xboxs, CPU/ GPU, Playtations, Rokus, TVs, receivers, mondems, routers, DVRs, window fans ,network appliances, DIY aquarium cooling and other audio video electronics
- Variable Speed Control: 110V - 220V Fan power supply with speed control function, turn the knob to adjust the speed, 4V - 12V adjustable fan speed，and can turn off the fan . | Input: 100V - 240V 50/60Hz | Output: DC 3-12V 200-2000ma
- DIY Vertical Window Fan: Can both vertical and horizontal, provide efficient cooling and ventilation. Mining rigs rely on the cooling power of fans for optimal operation.Double Metal Protective, the fan is equipped with double metal protective net
- Easy to Install: Draw out air in refrigerators, provide ventilation in greenhouses, prevent amplifier overheating, and vent hot air from living room consoles like PS4. Y cable connects 2 fans, two fans can be 42cm/16.5 in far away from each other
- Dual Ball Bearing: 240mm x 240mm x 25mm / 9.45in(L) x 4.72in(W) x 1in(H) in in total. | Rated Voltage :12V | Rated Current: 0.93A at full speed | Airflow: (82CFM)x4 at 12V | Speed: 2500 RPMx4
The three power values you must keep separate
Operators often compare the wrong ceiling. Depending on GPU generation, driver, and platform, telemetry can expose a requested limit, a current software limit, a default limit, and—on selected newer data-center modules—an enforced module limit. They are not interchangeable.
| Value | What it represents | Why it matters |
|---|---|---|
| Requested or configured limit | The ceiling requested through a supported management interface. | It shows operator intent, not necessarily what the hardware is allowed to consume. |
| Current software GPU limit | The active GPU-level limit reported by the driver. | Useful for checking whether a setting was accepted and is in force. |
| Default limit | The product’s documented or firmware-provided default. | Provides context when a host policy or administrator has changed the current value. |
| Enforced module limit | The limit actually selected by power-management policy on supported Hopper-and-newer data-center modules. | This can be lower than the software request. It is the critical comparison when available. |
On supported newer modules, NVIDIA documents the enforced limit as the minimum of several applicable limits. A host request can therefore appear valid while the module is still governed by a lower platform, firmware, or management constraint. Module-level fields are not available on every GPU or driver; check the installed version’s query help rather than assuming they exist.
Why a server may enforce a lower limit
GPU-local software power management
A driver or administrator can set a GPU power ceiling. When a compute workload would exceed it, the GPU adjusts its performance state. The resulting “SW Power Cap” reason means clocks are being optimized to remain within the currently set power limits. It does not identify who set the limit or establish that the value is wrong.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Module and platform policy
On supported products, the module can combine several ceilings before selecting the operative one. Firmware, host-side settings, and out-of-band management may each impose a limit. The most restrictive value wins, so changing only the visible nvidia-smi setting may have no effect.
Server-level power budgeting
NVIDIA’s DGX H100, DGX H200, and B200 documentation describes three relevant sources: VBIOS policy, host-side nvidia-smi configuration, and out-of-band SMBPBI controls. Their GPU Performance Monitoring Unit (PMU) “selects the most conservative policy to cap power consumption on a system.” This behavior is specific to the cited DGX platforms and should not be generalized to every OEM server. On another platform, the BMC, chassis controller, PSU redundancy mode, or virtualization layer may provide the equivalent constraint.
Electrical and cooling headroom
A server can intentionally reserve power for CPUs, fans, storage, networking, and redundant-PSU operation. Firmware may lower accelerator budgets when a supply is degraded, a redundant mode is selected, or inlet temperature rises. These are platform decisions, not necessarily GPU defects.
Rank #2
- An intelligent fan system designed for cooling audio video, DJ, server, network, and IT equipment racks.
- Protects rack-mount equipment from overheating, performance issues, and shortened lifespans.
- Programmable thermostat controller with automated speed control, alarm warnings, and backup memory.
- Premium anodized aluminum construction with CNC-machined detailing for a professional appearance.
- Size: 3U Rack Space | Design: Intake | Airflow: 60 to 300 CFM | Noise: 12 to 38 dBA | Bearings: Dual Ball
Throttle reasons that look like power throttling
Clock-limit fields are clues. Read them alongside temperature, power, clocks, utilization, and the behavior of the other GPUs in the same group.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors| Event or limiter | What it means | Next diagnostic focus |
|---|---|---|
| SW Power Cap | Clocks are optimized to stay within currently set power limits. | Compare draw with current and, where exposed, enforced limits; inspect host and platform policy. |
| SW Thermal | Software thermal control is protecting GPU or memory temperature limits. | Check temperature time series, airflow, fan behavior, heatsink contact, and inlet conditions. |
| HW Thermal | Hardware thermal slowdown is engaged. | Treat as an urgent cooling or environmental investigation; correlate with sensor readings. |
| HW Power Brake | An external power-brake signal is implicated, such as one asserted by system power delivery. | Investigate PSU, chassis power state, cabling, and platform management logs. |
| HW Slowdown | A broad hardware slowdown indication. NVIDIA lists high temperature, an external power brake, fast protection from excessive draw, and possible reporting during a P-state or clock transition. | Use time-series correlation; do not label it power throttling without supporting evidence. |
| Application clock setting | Configured application clocks bound the operating clocks. | Inspect administrator-set clock policies and reset or adjust them only through an authorized procedure. |
| Sync Boost | Every GPU in a synchronization group is held to the lowest possible clocks across that group. | Inspect all members of the group. A different GPU can be the one imposing the low clock. |
“Hardware slowdown” is intentionally broad. A transient clock or performance-state transition can set it without a persistent fault. The event list should therefore be interpreted with timestamps, not as a single universal error code.
Information to capture before changing anything
Start by recording the system identity. Include the exact GPU model and SKU, server platform, driver version, DCGM version, workload and software stack, and whether the accelerator is a physical GPU, partition, or virtualized device. Field names and available controls vary by product and release. Nominal wattage from a different SKU or server is not a valid comparison.
Define the symptom quantitatively. Record throughput (for example, inferences per second or training steps per second), latency, batch size, precision mode, and the time at which the slowdown starts. A workload that is input-bound or synchronization-bound may show low GPU power even when the application is performing normally.
Collect samples frequently enough to align telemetry with the symptom. At minimum, capture GPU utilization, SM and memory clocks where available, temperature, power draw, the relevant power-limit fields, and clock-event reasons. Preserve the same interval before, during, and after a slowdown so a transient event is distinguishable from a sustained cap.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Step-by-step diagnosis workflow
- Reproduce the slowdown. Run the representative AI workload with a fixed model, batch, precision, and concurrency. Note the healthy and degraded throughput or latency and the exact start time of each episode.
- Query supported power and performance data. A useful starting point is
nvidia-smi -q -d POWER,PERFORMANCE,CLOCK,TEMPERATURE. Review the output for current, requested, and default limits, power draw, temperatures, clocks, and performance-state information. Fields differ by generation. - Collect a compact time-series sample. For fields supported by the installed driver, try
nvidia-smi --query-gpu=name,power.draw,power.limit,clocks.current.sm,clocks.current.memory,temperature.gpu,utilization.gpu --format=csv. Repeat it at a regular interval or use a monitoring system. Do not treat one idle sample as evidence of throttling. - Check the query interface before relying on a field. Run
nvidia-smi --help-query-gpuand compare the listed names with your script. Module-power fields are available only on selected newer data-center products; a missing field is not itself a fault. - Read clock-event reasons during the event. Use
nvidia-smi -q -d PERFORMANCEor supported selective query fields. If DCGM is deployed, collect its clock-event-reason field as well. Confirm exact field names against the installed release documentation. - Compare draw with every applicable ceiling. If power draw approaches the current limit while clocks fall and SW Power Cap is asserted, a configured cap is a leading explanation. If an enforced module limit is exposed and lower than the request, investigate firmware, host, and out-of-band policy instead of repeatedly raising the request.
- Branch on the observed reason. Thermal events direct you toward temperature and cooling. HW Power Brake directs you toward external power delivery. Application-clock events direct you toward clock policy. Sync Boost requires inspection of every GPU in its group. HW Slowdown requires correlation with all of these signals.
- Inspect platform policy. On the documented DGX platforms, review VBIOS, host settings, SMBPBI controls, and PMU-selected policy using platform-specific procedures. On other servers, inspect the vendor’s BMC and power-budget logs.
- Use a controlled power test only when its prerequisites are met. DCGM’s
targeted_powertest sustains matrix-multiplication load on selected physical GPUs. Run it on GPUs that are idle before the test, with no competing workload. An enforced limit below the configured target can prevent the target from being reached. - Change a limit only after authority and headroom are proven. Confirm the effective ceiling, chassis and PSU budget, cooling capacity, redundancy policy, firmware support, and vendor operating constraints. Record the original setting and make one controlled change at a time.
Command starting points and how to read them
Detailed snapshot
nvidia-smi -q -d POWER,PERFORMANCE,CLOCK,TEMPERATURE is useful when you need the surrounding context rather than a single CSV row. Look for power draw and all reported limit values, then compare clocks and temperatures from the same snapshot. The command’s sections and fields vary by driver.
Rank #3
- [Adjustable] Adjustable temperature control helps ensure optimal performance for your rackmount such as network, server, music, and AV cabinets
- [Quiet and powerful] Equipped with three powerful 4” (120mm) noise control ball bearing fans capable of pumping 225 CFM of air, preventing overheating of expensive equipment
- [Optimal Airflow] This three fan cooling system will provide excellent cooling with its high-performance fans, which keep the hot air stream away from your setup with its top exhaust cool air system.
- [Compact Design] Device is standardized to mount to any 19" server rack or cabinet while taking only a single unit (1U) of space and has a wide variety of applications.
- [Programmable] Equipped with a programmable thermostat sensor controller for better temperature monitoring that will trigger fans based on your parameter configuration.
Script-friendly sampling
nvidia-smi --query-gpu=name,power.draw,power.limit,clocks.current.sm,clocks.current.memory,temperature.gpu,utilization.gpu --format=csv produces a compact record suitable for repeated collection. Add only fields confirmed by --help-query-gpu; unsupported fields can cause a query to fail or produce incomplete data.
Scrolling monitor
nvidia-smi dmon provides a scrolling view where supported. Use it to see whether power, utilization, clocks, and temperature change together. It is a live aid, not a replacement for saved, time-aligned logs.
DCGM event data
DCGM exposes clock-event reasons and diagnostic plugins. Keep the DCGM version with the incident record because field availability and plugin behavior are release-dependent.
Using DCGM’s targeted_power test correctly
The targeted_power diagnostic applies sustained matrix multiplication to selected physical GPUs. Its purpose is to exercise power behavior under a controlled load, not to publish a universal performance score. The documented plugin defaults include a 120.0-second duration and a 0.75 minimum target ratio. Those are test parameters, not measured speedups or guaranteed power draw for your system.
Prepare the test as follows:
- Select physical GPUs supported by the diagnostic; partitioned or virtualized devices may not be eligible.
- Stop or isolate competing jobs and verify the selected GPUs are idle.
- Record the configured target and the effective power limit before starting.
- Run the diagnostic with the version-matched DCGM configuration.
- Interpret an unattainable target as a configuration or enforced-limit finding, not automatically as a hardware failure.
A workload already using the GPU can distort the result, and a platform-enforced ceiling below the target can make reaching the target impossible by design. Preserve the plugin output together with the limit and policy data that were present during the run.
How to distinguish the main failure branches
Power cap is the primary limiter
You have a strong power-cap diagnosis when a demanding workload drives utilization high, clocks decrease, SW Power Cap is active, and draw stays near the effective ceiling. Next determine whether that ceiling is an intentional tenant or platform budget. A requested value higher than an enforced module value points to policy rather than a driver refusal.
Rank #4
- Adjustable temperature control helps ensure optimal performance for rackmount such as network, server, music, and AV cabinets
- Noise controlled fans makes the cooling system useful for a quiet office or business space
- Compact design mounts to any 19" inch cabinet and takes up only 1 unit of space
- Simple and easy to use LCD display allows user to control temperature
- Air pumped through to the top exhaust system of the fan
Temperature is the primary limiter
Thermal events should coincide with rising GPU or memory temperature and clock reduction. Check inlet temperature, airflow direction, fan control, dust or obstruction, heatsink and thermal-interface condition, and whether neighboring devices are recirculating hot exhaust. Do not raise a power cap while a thermal event is active; more electrical input can worsen the condition.
External power delivery is involved
HW Power Brake indicates an asserted external brake signal. Examine PSU health, chassis power mode, redundant operation, power cables and connectors, and management-controller logs. A GPU-local limit change cannot clear a chassis-level brake.
Application clocks are restricting performance
If the event identifies an application clock setting, inspect the administrator’s clock policy and any orchestration profile that reapplies it. Compare the configured clock bound with the clocks requested by the workload. Remove or adjust the policy only when you control the host and have validated the consequences.
Sync Boost is making one GPU look slow
With synchronization enabled, all GPUs in the group “will boost to the minimum possible clocks across the entire group.” The slow member may therefore be another GPU with a thermal, power, or policy issue. Collect event reasons and temperatures for every member before changing the apparently affected device.
The workload is not demanding the GPU
Low utilization, low power, and low clocks can be normal when the CPU, input pipeline, storage, network, kernel launch rate, or synchronization is the bottleneck. Compare GPU telemetry with application traces and host utilization. Do not call this power throttling unless the GPU is receiving enough work to reach the suspected ceiling.
Platform policy: what is specific to DGX systems
For DGX H100, H200, and B200, NVIDIA documents power-limit sources spanning VBIOS, host-side nvidia-smi, and out-of-band SMBPBI. The PMU selects the most conservative policy. Therefore, an operator can observe a lower effective limit even after setting a higher host request. Use the DGX release documentation and approved management workflow to identify which source is active.
Best Value
- A quiet fan kit designed for standard 19” racks, to be mounted on the roof or to replace existing fans.
- Features a speed controller utilizing PWM which can control the fan's speed without generating noise.
- Compatible with CLOUDPLATE series rack fans and can be linked to share the same programming.
- Heavy-Duty steel construction with spiral fan guards, mounting hardware, and power adapter.
- Size: Standard 120mm Rack Fans | Fans: 2 | Airflow 200 CFM | Noise: 26 dBA | Bearings: Dual Ball
Do not transfer this exact control path to an unrelated server. Other OEMs may expose different BMC names, firmware controls, or no user-adjustable module limit at all. Always report the platform model and software release when escalating the incident.
When changing a power limit is appropriate—and when it is not
Changing a limit can be reasonable only after you establish that the current ceiling is an intentional configuration choice, that the operator has authority to change it, and that the server has electrical and thermal headroom. Validate PSU capacity in the intended redundancy mode, rack and circuit limits, fan and cooling capacity, firmware support, and any service-level power budget.
Do not infer a safe wattage or expected speedup from the GPU model name alone. This evidence does not establish a universal threshold, performance-loss percentage, or benefit for an unspecified configuration. A higher cap can increase heat, fan power, rack consumption, and the chance of a platform power brake; it can also be ignored when a lower system policy remains enforced.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Make changes incrementally, document the old and new values, rerun the representative workload, and watch throughput, latency, temperature, power draw, event reasons, and neighboring devices. Restore the prior setting if thermal or platform-protection events appear.
Common diagnostic mistakes
- Using a single power reading. An instantaneous value taken during an input or synchronization stall says little about throttling.
- Assuming the requested limit is effective. On supported modules, a lower enforced limit can win.
- Comparing unlike GPUs by watts. Different SKUs and server designs have different voltage, cooling, and workload characteristics.
- Treating every clock drop as a power cap. Thermal, application-clock, sync-boost, power-brake, and broad hardware-slowdown events have different remedies.
- Testing a busy GPU with
targeted_power. Competing work invalidates the interpretation of the target result. - Raising the cap before checking cooling. This can turn a manageable limit into a thermal incident.
- Ignoring other GPUs. Sync Boost and shared chassis policies can make the root cause reside elsewhere.
- Assuming every field exists. Verify support with the installed
nvidia-smihelp and version documentation. - Blaming the GPU for a server brake. An external power signal requires platform and power-delivery investigation.
A compact decision framework
| Observation during a reproducible slowdown | Most useful next question |
|---|---|
| High utilization, falling clocks, SW Power Cap, draw near effective limit | Is the effective limit intentionally configured, or is a lower platform policy overriding the request? |
| Temperature reaches a limit with SW Thermal or HW Thermal | What cooling, airflow, inlet-temperature, or memory-temperature condition is triggering protection? |
| HW Power Brake appears | Which PSU, chassis, BMC, or external signal asserted the brake? |
| Application clock setting appears | Which host or orchestration policy imposed the clock bound? |
| Sync Boost appears | Which GPU in the synchronization group has the lowest permitted clocks and why? |
| HW Slowdown appears without a persistent cap | Did a transient thermal, power-brake, protection, or P-state transition coincide with the event? |
| Low power and low utilization with no limiter | Is the application waiting on CPU, data, storage, network, or synchronization rather than being power-limited? |
What a useful incident record contains
For support or capacity review, provide the GPU and server identity, driver and DCGM versions, physical versus partitioned or virtualized status, workload parameters, healthy and degraded performance, time-stamped power and clock samples, temperatures, all reported limit values, clock-event reasons, and any VBIOS, host, BMC, or SMBPBI policy settings. Include whether other GPUs changed behavior at the same time and whether the server was in a non-default redundancy or power mode.
That record lets an engineer separate a GPU-local software cap from a chassis policy, a thermal event, a workload bottleneck, or a transient protection transition. It also prevents an unsafe change based on an assumed wattage or an unverified field.
Bottom line
Power-limit throttling is the GPU doing what its power-management policy requires: reducing clocks to remain inside an enforced envelope. Diagnose it by reproducing the workload, aligning throughput with clocks, draw, temperatures, and event reasons, and comparing the requested ceiling with the actual enforced one. Only after platform authority, electrical budget, cooling, and policy constraints are clear should you consider changing a limit.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
