Coordinated load modulation of 1,000 GPU in the local power grid with a capacity of 1 MW with 90% DER - and total harmonic distortion of the current flies to 46.8%, and the damping coefficient fails to -0.27. The system is invaluable. This is a synchronized model of the worst case - but the numbers are not obtained on a napkin: impedance analysis, modeling of power systems, full-scale experiments on real GPUs and network inverters. Preprint of researchers from Zhejiang University (arXiv ID is not verified: the specified identifier 2607.05993 does not correspond to the YYMM.NNNNN format with a valid date) formalizes a new class of cyber-physical attacks - Bit2Watt: an attacker, working as a legitimate cloud tenant, rocks the GPU load to destabilize the power network of the data center. At the same time, no device on the side of the energy infrastructure is compromised. No Russian-language source did not disassemble this attack - below the first detailed analysis of mechanics, restrictions and countermeasures.
Business logic of attack: why modulate the power consumption of the GPU
Industroyer/CrashOverride (Ukraine, 2016), TRITON/Tris (Saudi Arabia, 2017) - the classic cyber attacks on energy - a detailed analysis of such vectors is collected in full power network protection card and OT systems. Both demanded compromise of OT devices: RTU controllers, HMI panels, SCADA-servers. The attacker overcame the perimeter, gained access to the industrial network, understood the protocols. The Bit2Watt attack works differently: the impact goes not through a network protocol, but through physics - through controlled jumps in the power consumption of computing equipment.
Why does this attacker? There are several options. Sabotage of competing infrastructure: cascading failure of the power system interrupts the training of models, destroys the SLA. Hidden denial-of-service: UPS bypass, circuit breakers, ATS - computational load reset. Hours-long training checkpoint rolls back, financial losses are measured in hundreds of thousands of GPU hours. The data center’s energy system is part of the infrastructure, and the number of cloud intrusion cases increased by 26% year-on-year (CrowdStrike Global Threat Report 2025). The surface for attacks that require only cloud access is expanding.
And here’s what’s unpleasant: Bit2Watt leaves no traces at the file system level, does not use malware, does not require an increase in privileges. The attacker works through the regular API of the cloud platform - CUDA Runtime API, PyTorch distributed training, standard job scheduler. Cyber-physical attacks on data centers of this class do not fall under any existing signature detect.
The Bit2Watt threat model: a cloud tenant as a cyber-physical weapon
According to the preprint Zhou et al. (arXiv ID is not verified), the enemy rents GPU power through the standard interface of the cloud provider. No software vulnerabilities, no lateral movement, no BMC/IPMI compromise. Instead, are a legitimate user-level API to run a CUDA core or training pypalines structured so that the power consumption pattern of the GPU-cluster puts controlled disturbances in the data center power grid.
This is a fundamentally new attack surface for the safety of data center power supply. Cloud monitoring monitors the utilization of GPU, network traffic, memory capacity - but does not analyze the frequency spectrum of the power oscillations caused by calculations. Facility-monitoring (DCIM platforms such as Nlyte or Sunbird, PDU-Teletra Lighting via SNMP) captures weighted life intake, but the standard frequency of the SNMP-based PDU survey from Raritan, APC, Vertiv - 1-5 seconds. This is three or four orders of coarsely crudely than the frequency of SWMA modulation (up to 6 000 Hz according to the authors' stated data). Between the cloud-side and-side monitoring facility is a dead zone. This is where Bit2Watt works.
The context of scale: the projected growth in electricity consumption by data centers is from 415 TWh in 2024 to more than 945 TWh by 2030 (OPAL-RT data with reference to Goldman Sachs Research). The volume of computing power available to the attacker will only grow.
A place in the attack chain
In terms of MITRE ATT&CK Enterprise matrix The Bit2Watt chain looks like this:
Discovery - System Information Discovery (T1082): the attacker determines the topology of the GPU-cluster, the number of SM (streaming multiprocessors), the clock frequencies, the GPU model. Through standard CUDA API calls or nvidia-smi- it's nothing unusual.
Execution is the launch of a CUDA core through a legitimate user-level API. Formally, a separate ATT&C-technics is not required: CUDA-code is the standard function of the leaseable resource. When using Python trouts (PyTorch, TensorFlow) is formally applicable T1059.006 - Python, although in this context it is a legitimate code.
Impact - Result: Compute Hijacking (T1496.001) - the use of resources for the tasks of the attacker; Service Stop (T1489) - cascade failure when triggering protections; Runtime Data Manipulation (T1565.003) - structured manipulation of the pattern of execution of computations for cyber-physical effect.
In the context of the OT-infrastructure of energy supply, the consequences are conceptually consistent with the categories ATT&CK for ICS matrix (v14) - a denial of control loss and of power control system. But Bit2Watt crosses the IT/OT border without penetration into the OT circuit: the impact is transmitted through a physical channel (electric grid), rather than through a network connection. This makes the attack really interesting in terms of research.
LTMA: manipulation of GPU load under the LLM training mask
LTMA (LLM Training Modulation Attack) is a more sophisticated vector. Instead of clearly alternating the load, the attacker embeds modulation logic in the standard pipelline of neural network training. Distributed training naturally creates oscillations: the forward/backward pass (forward/backward pass) is replaced by the phase of synchronization of gradients (all-reduce), where the GPU is waiting for data from the network. This pattern is also described in Microsoft Research’s “Power Stabilization for AI Training Datacenters” study: tens of thousands of GPUs in a synchronous training mission create significant power fluctuations.
The attacker enhances the natural pattern, selecting the size of the mini-fuck, the number of GPUs in the distributed training and the synchronization parameters so that the resulting frequency of power oscillations falls into the resonant range of the energy infrastructure. From the point of view of the cloud provider, the load looks like a normal training - neither anomalous system calls, nor suspicious recycling patterns. LTMA’s inolable visibility is significantly higher than that of SWMA: the legitimate training workload is indistinguishable from an attack without analyzing the frequency spectrum of energy consumption.
Environment requirements for both vectors: a minimum of 2 GPU (NVIDIA Volta and newer, compute >= 7.0, including Turing/Ampere/Hoper) with support __nanosleep, entered in PTX ISA 6.3 / CUDA 10 for sm_70+; CUDA Toolkit >= 10.0 (recommended for Appere/Hopper); for LTMA additional PyTorch >= 2.0 with NCCL backend; monitoring consumption through LTMA nvidia-smi --query-gpu=power.draw --format=csv -l 1 or NVML API.
Destabilization of the data center power grid: from harmonics to cascade failure
To understand how computing modulation is broadcast to physical damage, it is worthwhile to trace the current from utility to GPU.
Electricity from grid comes through the transformer substation, passes through the double conversion UPS (AC-DC-AC), distributed through the PDU on racks. Each GPU server contains pulsed power supplies (PSU) - the final AC-DC conversion. In parallel, more and more data center is using on-site renewable energy - solar panels and wind generators connected through inverters. According to OPAL-RT, DER-inverters operate on frequencies critical of grid stability, and their interaction with nonlinear loads of GPU clusters creates the potential for resonance.
According to impedance analysis in the preprint Zhou et al. (arXiv ID is not verified), at a high proportion of DER (90%) in the local data center network, the load modulation of the GPU-cluster load modulation interacts with inverter control loops. The stated results for the synchronized model of the worst case: with the modulation of 1 000 GPU in the 1 MW current THD network reaches 46.8% - RMS of harmonic components of the current is ~ 47% of the fundamental. This increases losses I^2R in conductors by about 22% and causes additional heating of transformers. The damping coefficient drops to -0.27 - unstable fashion.
Chain of physical consequences: increased THD -> overheating of transformers and neutral conductors -> negative damping -> voltage autosaging -> output of parameters for admission of IEEE 519 -> triggering of protective machines -> load reset -> reset -> repayment of inverse power. In an extreme simulated scenario - cascade failure in the distribution network. The authors note that a blackout can affect more than 80% of the transmission system, but they emphasize: this is a simulated worst case, not a field experiment.
How OT-infrastructure data centers differs from IT environment
The data center power system is the OT-environment with all the ensuing restrictions. Power management systems use Modbus RTU/TCP: FC03 (Read Holding Registers) to read voltage, current and power readings; FC06/FC16 (Write Single/Multiple Registers) to control the ATS switching thresholds and UPS settings. SNMP v2c/v3 - for a PDU survey. DCIM platforms aggregate telemetry, but the typical frequency of the survey is seconds, not milliseconds.
In IT-infrastructure patching - routine. OT is a different story. The average time between CVE posting and elimination is 29 months (IBM X-Force 2025). For OT equipment, this period is often even longer. An indicative example – CVE-2024-2617 in Hitachi RTU500 controllers (property version series 12.6.6.0–13.5.2.0): bypassing the secure update mechanism allows you to download unsigned firmware if there is authenticated with high privileges (CVSS 7.2, HIG; vector CVSS:3.1/AV:N/AV/C/L:L/L/UI:C/C/H/C/I/I). Authenticized access to OT equipment - and you can pour unsigned firmware through the standard interfaces.
Standard IT-means of protection - EDR, SIEM, network firewalls - do not see high-frequency power fluctuations on the power bus. OT-monitoring, even through the Modbus FC03, a survey of registers, operates on frequencies insufficient to detect SWMA-modulation. UPS destabilization of the data center through workload-modulation falls into the dead zone between IT and OT monitoring. This is a fundamental problem of data center power supply.
Watt2Bit Feedback: Disability of Computation and Data Leakage
The authors describe the potential feedback mechanism - Watt2Bit. If the load modulation causes voltage fluctuations on the power bus, these oscillations hit the computing equipment in the same data center.
Voltage jumps lead to errors in GPU memory (ECC-mistakes, bit-flips), interrupt model training, start restarting servers and moving UPS to bypass mode. For large AI training, the loss of synchronization of even one GPU can roll back tens of minutes of training. In cascade failure - hours of computational time and loss of checkpoint data. It turns out a self-suffocating loop: initial modulation causes instability, instability causes additional load resets, resets create new jumps - the cycle increases before the triggering of emergency protection.
The second aspect is covert exfiltration through EMI side channel. Current fluctuations create electromagnetic radiation, which theoretically can be left out of the data center. The channel is extremely low-speed, but for the ex-filtering of cryptographic keys or short tokens, it may be enough. GPU energy consumption is turning into a dual-purpose vulnerability: both the attack vector on the infrastructure and a potential leakage channel.
Restrictions of Bit2Watt attacks and effectiveness factors
The results of Zhou et al. are obtained in a synchronized model of the worst case. Between the model and reality - the distance, and it is worth being evaluated soberly.
Synchronization. For maximum effect, all modulated GPUs must switch the load shinfas. In a distributed cloud environment, the timing jitter between the nodes (tens to hundreds of microseconds at the level of NIC-to-NIC latency) partially desynesy synchronizes modulation and reduces the amplitude of oscillations. The authors analyze the effect of jitter (Appendix C preprint), but the quantitative decline in efficiency depends on the specific network infrastructure and placement topology.
Share of DER. THD 46.8% and negative damping are obtained at 90% DER - a scenario not yet typical for most data centers. When dominating utility grid (synchronous generators) damping the system is much higher, and the modulation effect weakens. But the trend to increase on-site solar and battery storage in the data center of hyperscalers is moving precisely to an increase in DER-shadoles.
The scale of available resources. 1 000 GPUs in one 1-MW segment - several dozen racks with a load of 50-70 kW on the rack for AI-clusters. The attacker needs to rent a significant share of the power of one physical segment of the power grid. Policy of placing a provider (tantant isolation, spread placement groups) can distribute the GPU attacker to different PDU feeders, reducing the coherence of the impact.
Warp scheduling. Mechanism __nanosleep It works at the warp level. If the GPU simultaneously spins the tasks of other tenants (MIG partition, time-sliced sharing), the planner will fill the vacant SM with someone else's warps. In shared-tenancy environments, the effectiveness of SWMA degrades in proportion to the volume of extraneous load.
Evolution of Monitoring. The standard cloud-side monitoring of Bit2Watt does not detect, but some providers begin to introduce power telemthry at the GPU level through the NVML API. Abnormal periodic pattern power.draw with a stable frequency > 10 Hz on the GPU group - a potential IoC for this class of attacks.
Countermeasures: cross-layer protection of data center power supply
BitWatt authors directly point to the need for cross-layer defenses - protective measures that combine the cloud, and grid levels. Below is a checklist for transmission to engineering teams.
Cloud level (Compute/Cloud):
Implement monitoring of the frequency spectrum of GPU power consumption via NVML (nvmlDeviceGetPowerUsage) with sampling at least 100 Hz - standard DCIM does not provide
Configure the detective rule: periodic oscillations of power with a stable frequency > 10 Hz on cluster > 100 GPU - SWMA indicator
Enter a placement policy: prevent co-location GPU from one tenant > 50% of the power of one PDU segment (at orchestrator / placement groups level)
Activate GPU Power Cappping (nvidia-smi -pl <watts>) to limit the amplitude of possible modulation of the meam on clusters
Level (DCIM/Power):
Install power analyze qualityr at the input of each PDU feeder with a sampling frequency >= 10 kHz (Dranetz, Fluke 1770 or analogues)
Configure alerates to THD voltage > 5% by PDU feeder for <1 kV systems (IEE 519-2014, Table 1) and TDD current > 5-20% depending on Isc/IL ratio (IEEE 519-2014, Table 2)
Monitor power factor in real time: PF drop below 0.9 with a stable declared load - harmonic distortion indicator
Level of power grid (Grid/DER):
For data center with DER > 30%: conduct impedance scan in the range of 100 Hz - 10 kHz to detect resonant frequencies of the local network
Implement active harmonic filtering on DER inverters
Consider review the energy battery storage systems (BESS) as a buffer between DER and load - smoothes high-frequency oscillations
Business logic of attack: why modulate the power consumption of the GPU
Industroyer/CrashOverride (Ukraine, 2016), TRITON/Tris (Saudi Arabia, 2017) - the classic cyber attacks on energy - a detailed analysis of such vectors is collected in full power network protection card and OT systems. Both demanded compromise of OT devices: RTU controllers, HMI panels, SCADA-servers. The attacker overcame the perimeter, gained access to the industrial network, understood the protocols. The Bit2Watt attack works differently: the impact goes not through a network protocol, but through physics - through controlled jumps in the power consumption of computing equipment.
Why does this attacker? There are several options. Sabotage of competing infrastructure: cascading failure of the power system interrupts the training of models, destroys the SLA. Hidden denial-of-service: UPS bypass, circuit breakers, ATS - computational load reset. Hours-long training checkpoint rolls back, financial losses are measured in hundreds of thousands of GPU hours. The data center’s energy system is part of the infrastructure, and the number of cloud intrusion cases increased by 26% year-on-year (CrowdStrike Global Threat Report 2025). The surface for attacks that require only cloud access is expanding.
And here’s what’s unpleasant: Bit2Watt leaves no traces at the file system level, does not use malware, does not require an increase in privileges. The attacker works through the regular API of the cloud platform - CUDA Runtime API, PyTorch distributed training, standard job scheduler. Cyber-physical attacks on data centers of this class do not fall under any existing signature detect.
The Bit2Watt threat model: a cloud tenant as a cyber-physical weapon
According to the preprint Zhou et al. (arXiv ID is not verified), the enemy rents GPU power through the standard interface of the cloud provider. No software vulnerabilities, no lateral movement, no BMC/IPMI compromise. Instead, are a legitimate user-level API to run a CUDA core or training pypalines structured so that the power consumption pattern of the GPU-cluster puts controlled disturbances in the data center power grid.
This is a fundamentally new attack surface for the safety of data center power supply. Cloud monitoring monitors the utilization of GPU, network traffic, memory capacity - but does not analyze the frequency spectrum of the power oscillations caused by calculations. Facility-monitoring (DCIM platforms such as Nlyte or Sunbird, PDU-Teletra Lighting via SNMP) captures weighted life intake, but the standard frequency of the SNMP-based PDU survey from Raritan, APC, Vertiv - 1-5 seconds. This is three or four orders of coarsely crudely than the frequency of SWMA modulation (up to 6 000 Hz according to the authors' stated data). Between the cloud-side and-side monitoring facility is a dead zone. This is where Bit2Watt works.
The context of scale: the projected growth in electricity consumption by data centers is from 415 TWh in 2024 to more than 945 TWh by 2030 (OPAL-RT data with reference to Goldman Sachs Research). The volume of computing power available to the attacker will only grow.
A place in the attack chain
In terms of MITRE ATT&CK Enterprise matrix The Bit2Watt chain looks like this:
Discovery - System Information Discovery (T1082): the attacker determines the topology of the GPU-cluster, the number of SM (streaming multiprocessors), the clock frequencies, the GPU model. Through standard CUDA API calls or nvidia-smi- it's nothing unusual.
Execution is the launch of a CUDA core through a legitimate user-level API. Formally, a separate ATT&C-technics is not required: CUDA-code is the standard function of the leaseable resource. When using Python trouts (PyTorch, TensorFlow) is formally applicable T1059.006 - Python, although in this context it is a legitimate code.
Impact - Result: Compute Hijacking (T1496.001) - the use of resources for the tasks of the attacker; Service Stop (T1489) - cascade failure when triggering protections; Runtime Data Manipulation (T1565.003) - structured manipulation of the pattern of execution of computations for cyber-physical effect.
In the context of the OT-infrastructure of energy supply, the consequences are conceptually consistent with the categories ATT&CK for ICS matrix (v14) - a denial of control loss and of power control system. But Bit2Watt crosses the IT/OT border without penetration into the OT circuit: the impact is transmitted through a physical channel (electric grid), rather than through a network connection. This makes the attack really interesting in terms of research.
LTMA: manipulation of GPU load under the LLM training mask
LTMA (LLM Training Modulation Attack) is a more sophisticated vector. Instead of clearly alternating the load, the attacker embeds modulation logic in the standard pipelline of neural network training. Distributed training naturally creates oscillations: the forward/backward pass (forward/backward pass) is replaced by the phase of synchronization of gradients (all-reduce), where the GPU is waiting for data from the network. This pattern is also described in Microsoft Research’s “Power Stabilization for AI Training Datacenters” study: tens of thousands of GPUs in a synchronous training mission create significant power fluctuations.
The attacker enhances the natural pattern, selecting the size of the mini-fuck, the number of GPUs in the distributed training and the synchronization parameters so that the resulting frequency of power oscillations falls into the resonant range of the energy infrastructure. From the point of view of the cloud provider, the load looks like a normal training - neither anomalous system calls, nor suspicious recycling patterns. LTMA’s inolable visibility is significantly higher than that of SWMA: the legitimate training workload is indistinguishable from an attack without analyzing the frequency spectrum of energy consumption.
Environment requirements for both vectors: a minimum of 2 GPU (NVIDIA Volta and newer, compute >= 7.0, including Turing/Ampere/Hoper) with support __nanosleep, entered in PTX ISA 6.3 / CUDA 10 for sm_70+; CUDA Toolkit >= 10.0 (recommended for Appere/Hopper); for LTMA additional PyTorch >= 2.0 with NCCL backend; monitoring consumption through LTMA nvidia-smi --query-gpu=power.draw --format=csv -l 1 or NVML API.
Destabilization of the data center power grid: from harmonics to cascade failure
To understand how computing modulation is broadcast to physical damage, it is worthwhile to trace the current from utility to GPU.
Electricity from grid comes through the transformer substation, passes through the double conversion UPS (AC-DC-AC), distributed through the PDU on racks. Each GPU server contains pulsed power supplies (PSU) - the final AC-DC conversion. In parallel, more and more data center is using on-site renewable energy - solar panels and wind generators connected through inverters. According to OPAL-RT, DER-inverters operate on frequencies critical of grid stability, and their interaction with nonlinear loads of GPU clusters creates the potential for resonance.
According to impedance analysis in the preprint Zhou et al. (arXiv ID is not verified), at a high proportion of DER (90%) in the local data center network, the load modulation of the GPU-cluster load modulation interacts with inverter control loops. The stated results for the synchronized model of the worst case: with the modulation of 1 000 GPU in the 1 MW current THD network reaches 46.8% - RMS of harmonic components of the current is ~ 47% of the fundamental. This increases losses I^2R in conductors by about 22% and causes additional heating of transformers. The damping coefficient drops to -0.27 - unstable fashion.
Chain of physical consequences: increased THD -> overheating of transformers and neutral conductors -> negative damping -> voltage autosaging -> output of parameters for admission of IEEE 519 -> triggering of protective machines -> load reset -> reset -> repayment of inverse power. In an extreme simulated scenario - cascade failure in the distribution network. The authors note that a blackout can affect more than 80% of the transmission system, but they emphasize: this is a simulated worst case, not a field experiment.
How OT-infrastructure data centers differs from IT environment
The data center power system is the OT-environment with all the ensuing restrictions. Power management systems use Modbus RTU/TCP: FC03 (Read Holding Registers) to read voltage, current and power readings; FC06/FC16 (Write Single/Multiple Registers) to control the ATS switching thresholds and UPS settings. SNMP v2c/v3 - for a PDU survey. DCIM platforms aggregate telemetry, but the typical frequency of the survey is seconds, not milliseconds.
In IT-infrastructure patching - routine. OT is a different story. The average time between CVE posting and elimination is 29 months (IBM X-Force 2025). For OT equipment, this period is often even longer. An indicative example – CVE-2024-2617 in Hitachi RTU500 controllers (property version series 12.6.6.0–13.5.2.0): bypassing the secure update mechanism allows you to download unsigned firmware if there is authenticated with high privileges (CVSS 7.2, HIG; vector CVSS:3.1/AV:N/AV/C/L:L/L/UI:C/C/H/C/I/I). Authenticized access to OT equipment - and you can pour unsigned firmware through the standard interfaces.
Standard IT-means of protection - EDR, SIEM, network firewalls - do not see high-frequency power fluctuations on the power bus. OT-monitoring, even through the Modbus FC03, a survey of registers, operates on frequencies insufficient to detect SWMA-modulation. UPS destabilization of the data center through workload-modulation falls into the dead zone between IT and OT monitoring. This is a fundamental problem of data center power supply.
Watt2Bit Feedback: Disability of Computation and Data Leakage
The authors describe the potential feedback mechanism - Watt2Bit. If the load modulation causes voltage fluctuations on the power bus, these oscillations hit the computing equipment in the same data center.
Voltage jumps lead to errors in GPU memory (ECC-mistakes, bit-flips), interrupt model training, start restarting servers and moving UPS to bypass mode. For large AI training, the loss of synchronization of even one GPU can roll back tens of minutes of training. In cascade failure - hours of computational time and loss of checkpoint data. It turns out a self-suffocating loop: initial modulation causes instability, instability causes additional load resets, resets create new jumps - the cycle increases before the triggering of emergency protection.
The second aspect is covert exfiltration through EMI side channel. Current fluctuations create electromagnetic radiation, which theoretically can be left out of the data center. The channel is extremely low-speed, but for the ex-filtering of cryptographic keys or short tokens, it may be enough. GPU energy consumption is turning into a dual-purpose vulnerability: both the attack vector on the infrastructure and a potential leakage channel.
Restrictions of Bit2Watt attacks and effectiveness factors
The results of Zhou et al. are obtained in a synchronized model of the worst case. Between the model and reality - the distance, and it is worth being evaluated soberly.
Synchronization. For maximum effect, all modulated GPUs must switch the load shinfas. In a distributed cloud environment, the timing jitter between the nodes (tens to hundreds of microseconds at the level of NIC-to-NIC latency) partially desynesy synchronizes modulation and reduces the amplitude of oscillations. The authors analyze the effect of jitter (Appendix C preprint), but the quantitative decline in efficiency depends on the specific network infrastructure and placement topology.
Share of DER. THD 46.8% and negative damping are obtained at 90% DER - a scenario not yet typical for most data centers. When dominating utility grid (synchronous generators) damping the system is much higher, and the modulation effect weakens. But the trend to increase on-site solar and battery storage in the data center of hyperscalers is moving precisely to an increase in DER-shadoles.
The scale of available resources. 1 000 GPUs in one 1-MW segment - several dozen racks with a load of 50-70 kW on the rack for AI-clusters. The attacker needs to rent a significant share of the power of one physical segment of the power grid. Policy of placing a provider (tantant isolation, spread placement groups) can distribute the GPU attacker to different PDU feeders, reducing the coherence of the impact.
Warp scheduling. Mechanism __nanosleep It works at the warp level. If the GPU simultaneously spins the tasks of other tenants (MIG partition, time-sliced sharing), the planner will fill the vacant SM with someone else's warps. In shared-tenancy environments, the effectiveness of SWMA degrades in proportion to the volume of extraneous load.
Evolution of Monitoring. The standard cloud-side monitoring of Bit2Watt does not detect, but some providers begin to introduce power telemthry at the GPU level through the NVML API. Abnormal periodic pattern power.draw with a stable frequency > 10 Hz on the GPU group - a potential IoC for this class of attacks.
Countermeasures: cross-layer protection of data center power supply
BitWatt authors directly point to the need for cross-layer defenses - protective measures that combine the cloud, and grid levels. Below is a checklist for transmission to engineering teams.
Cloud level (Compute/Cloud):
Implement monitoring of the frequency spectrum of GPU power consumption via NVML (nvmlDeviceGetPowerUsage) with sampling at least 100 Hz - standard DCIM does not provide
Configure the detective rule: periodic oscillations of power with a stable frequency > 10 Hz on cluster > 100 GPU - SWMA indicator
Enter a placement policy: prevent co-location GPU from one tenant > 50% of the power of one PDU segment (at orchestrator / placement groups level)
Activate GPU Power Cappping (nvidia-smi -pl <watts>) to limit the amplitude of possible modulation of the meam on clusters
Level (DCIM/Power):
Install power analyze qualityr at the input of each PDU feeder with a sampling frequency >= 10 kHz (Dranetz, Fluke 1770 or analogues)
Configure alerates to THD voltage > 5% by PDU feeder for <1 kV systems (IEE 519-2014, Table 1) and TDD current > 5-20% depending on Isc/IL ratio (IEEE 519-2014, Table 2)
Monitor power factor in real time: PF drop below 0.9 with a stable declared load - harmonic distortion indicator
Level of power grid (Grid/DER):
For data center with DER > 30%: conduct impedance scan in the range of 100 Hz - 10 kHz to detect resonant frequencies of the local network
Implement active harmonic filtering on DER inverters
Consider review the energy battery storage systems (BESS) as a buffer between DER and load - smoothes high-frequency oscillations
