Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures

Zhouhao Ji, Kaikai Pan, Wenyuan XuarXiv 2026 · arXiv

TL;DR

The paper uncovers Bit2Watt, a novel cyber-physical vulnerability where a standard cloud tenant can manipulate GPU workloads to induce high-frequency power fluctuations. These fluctuations exploit the electrical characteristics of data center power infrastructure and modern renewable grids, causing severe power-quality degradation, system instability, and potentially cascading grid failures.

What Problem Does This Paper Address?

As data centers scale up to support massive AI and high-performance computing (HPC) workloads, their power consumption has skyrocketed. To stay sustainable, these facilities are increasingly integrating on-site renewable energy resources like solar and wind. This creates a tightly coupled cyber-physical system where computational workloads directly influence the physical power grid.

Traditional cyber-physical attacks on power grids target grid-side control software or communication channels. This paper addresses a completely different, overlooked threat vector: Can a completely legitimate cloud tenant, using standard user-level programming privileges, weaponize computational workloads to destabilize the underlying power infrastructure?

The authors demonstrate that because modern grids are increasingly dominated by inverter-based resources (which lack the physical inertia of traditional power plants), they are highly sensitive to fast, high-frequency load variations that standard cloud monitoring tools fail to detect.

Overview of the Bit2Watt and Watt2Bit closed-loop risk.

Methods

The authors design and validate two distinct user-space power modulation techniques that require no elevated system privileges:

  1. Synthetic Workload Modulation Attack (SWMA): A custom CUDA program that rapidly toggles GPU execution between an active mode (high-intensity tensor-core or FP32 loops that maximize power draw) and a passive mode (idle state). A CPU controller uses unified memory (cudaMallocManaged) to poll and instantly switch modes, achieving precise, high-frequency power oscillations.
  2. LLM Training Modulation Attack (LTMA): A stealthier approach that embeds modulation logic directly into standard deep learning pipelines (e.g., PyTorch or DeepSpeed). By dynamically altering hyperparameters or inserting auxiliary operations during training iterations, the attack camouflages its malicious power signature as legitimate computational noise.

To understand how these workloads interact with the grid, the authors establish an impedance-based analytical model. This model treats the GPU cluster as a Constant Power Load (CPL) and maps its interaction with the frequency-dependent dynamics of inverter-based distributed energy resources (DERs).

Datasets

While this is primarily a security and systems paper rather than a traditional machine learning paper, the authors utilized specific workloads and grid models for evaluation:

  • Workload Data: The LTMA was evaluated using the training process of GPT-2. Baseline single-GPU power dynamics were contextualized using the MIT Supercloud Dataset (specifically tracking NVIDIA V100 power draw).
  • Grid Simulation Models: System-level impacts were evaluated using standard power system benchmarks, including a local 1-MW power system model and the wide-area IEEE 118-bus system, alongside a modified 9241-bus European transmission network to simulate large-scale cascading failures.

Key Insights and Ideas

  • The CPL Positive Feedback Loop: At sub-millisecond timescales, GPUs act as Constant Power Loads (CPLs) exhibiting Negative Incremental Resistance (NIR). If the grid voltage drops, the GPU's internal voltage regulator module automatically draws more current to maintain constant power. This creates a dangerous positive feedback loop that amplifies voltage fluctuations and degrades system damping.
  • Exploiting Inverter Resonances: Unlike traditional household loads that fluctuate at a few Hertz, GPU clusters can modulate power at frequencies exceeding 6,000 Hz. Modern inverter-dominated grids have high-frequency resonant peaks that these rapid modulations can directly excite.
  • The Telemetry Blind Spot: Standard cloud- and facility-side monitoring tools (like NVML, RAPL, or rack-level PDU telemetry) sample at coarse intervals (1 Hz to 1 kHz). Because the Bit2Watt attack concentrates its energy in high-frequency components, it remains entirely invisible to routine operational logging.
  • The Watt2Bit Feedback Loop: The attack doesn't just damage the grid; it loops back to the data center. High-frequency harmonics increase the Equivalent Series Resistance (ESR) in power delivery components, causing localized overheating and triggering hardware protection mechanisms that interrupt computing services (Denial of Service).

Key Results

  • High-Frequency Modulation: In real-world hardware tests across various NVIDIA GPUs (RTX 3090, RTX 4090, A100, etc.), SWMA achieved precise power modulation frequencies up to 6,000 Hz, while the stealthier LTMA reached up to 3,000 Hz.
  • Grid Destabilization: In a simulated 1-MW local grid with 90% renewable penetration, synchronizing just 1,000 GPUs spiked the Total Harmonic Distortion (THD) to 46.8% (far exceeding the 13% standard limit) and pushed the system damping ratio to -0.27, introducing severe instability.
  • Cascading Failures: When scaling the attack to a simulated 9241-bus European transmission network, a localized Bit2Watt disturbance affecting just 2% of the total system load triggered a massive cascade, disconnecting 1,238 transmission lines and causing an 81% total load loss.
  • Covert Data Exfiltration: By encoding binary data via Frequency-Shift Keying (FSK) onto the GPU workload (switching between 2 kHz and 200 Hz), the authors successfully exfiltrated data through electromagnetic interference (EMI) side channels with >99% accuracy.

Limitations and Future Work

The primary limitation of the worst-case attack scenario is its reliance on perfect synchronization across thousands of distributed GPUs. In reality, network latency and operating system scheduling jitter (timing jitter) introduce asynchrony. The authors note that a timing jitter of $\sigma_\tau = 100\ \mu\text{s}$ reduces the aggregate attack amplitude by roughly 20%, though an adversary could potentially compensate for this by scaling up the number of compromised tenant accounts.

Future work highlights the urgent need for cross-layer defenses. Security paradigms must evolve past isolated boundaries: cloud providers need finer-grained, profiling-aware workload monitoring, while data center power infrastructure requires active filtering and local energy buffering (like supercapacitors or battery storage) designed to absorb high-frequency transients.

Summary updated

Abstract

Modern data centers increasingly rely on large-scale GPU clusters and on-site renewable energy resources, resulting in a tightly coupled cyber-physical system between computing workloads and power-electronic-dominated grids. In this paper, we reveal Bit2Watt, a previously unexplored vulnerability in which an adversary manipulates GPU workloads to induce controlled, high-frequency power modulations that destabilize local power infrastructure and propagate back to disrupt computing services. Unlike traditional attacks that compromise grid-side devices or communication channels, Bit2Watt operates entirely within the cyber layer as a legal tenant, which could amplify fluctuations, harmonic distortion, and damping degradation, particularly in high-DER-penetration scenarios. This risk is difficult to detect under routine cloud- and facility-side monitoring because it exploits legitimate workload execution paths and concentrates much of its distinctive behavior in high-frequency components that are weakly captured by common telemetry. We validate Bit2Watt through impedance-based analysis, power system simulations, and real-world experiments on GPUs and grid-connected PV inverters. Under the synchronized worst-case aggregation model studied in the paper, manipulating 1,000 GPUs in a 1-MW local power system with 90% DERs raises current THD to 46.8% and results in a damping ratio of -0.27. We further show that the resulting power-quality degradation can stress data-center power-delivery equipment, trigger protection mechanisms, and, in extreme simulated cases, induce cascading failures in transmission-scale systems. In addition, we analyze a plausible Watt2Bit feedback path, including denial-of-service risks and covert information exfiltration via EMI side channels. This work highlights the urgent need for cross-layer defenses that jointly consider workload scheduling and power electronics.
BibTeX
@misc{ji2026bit,
  author = {Zhouhao Ji and Kaikai Pan and Wenyuan Xu},
  title = {{Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures}},
  year = {2026},
  eprint = {2607.05993},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.05993},
}
Figure