Book a demo

Cut patent&paper research from weeks to hours with PatSnap Eureka AI!

Try now

Low-Noise PDN Design for AI Accelerator Chips — PatSnap Eureka

Low-Noise PDN Design for AI Accelerator Chips — PatSnap Eureka
Tools Explore in Eureka
Reading14 min
PublishedJun 2025
Coverage2009–2025
PDN Design · AI Chips · 2025

Low-Noise Power Delivery Networks for AI Accelerator Chips

As voltage supply rails shrink and current demands surge with trillion-parameter workloads, supply noise directly degrades timing margins, bit error rates, and model inference accuracy. This landscape surveys patent and literature signals covering PDN noise reduction, on-chip power management, and AI-assisted optimization for neural network accelerator hardware.

Fig. 01 — PDN Patent Records by Technology Cluster (2009–2025)
PDN Patent Records by Technology Cluster: Hierarchical Impedance Simulation 3, Active/AI-Driven PDN Control 4, DVFS and Workload-Aware Management 3, Near-Threshold Voltage Design 3 Bar chart showing the distribution of patent and literature records across four PDN technology clusters for AI accelerator chips identified in the 2009–2025 dataset. Source: PatSnap Eureka.
Published by PatSnap Insights Team · · 14 min read Verified by PatSnap Eureka Data
Technology Overview

Three Interconnected Problem Domains in PDN Design

Low-noise PDN design for AI accelerators spans voltage droop control, frequency-domain impedance management, and adaptive real-time control loops — each compounding the others as process nodes shrink.

Power delivery network design for AI accelerator chips addresses one of the most critical constraints in modern compute silicon. The three core problem domains are: (1) controlling voltage droop and ripple caused by the massive, correlated current transients generated by synchronous matrix-multiplication engines; (2) managing impedance across the frequency spectrum from DC to multi-GHz switching harmonics; and (3) deploying active or AI-driven control loops that adapt PDN parameters in real time.

Across the retrieved dataset, publication dates span from 2009 to late 2025, with a clear inflection point around 2019–2021 driven by the rapid scaling of DNN workloads and the simultaneous push below 10 nm process nodes. Research from organisations including NVIDIA, Amazon Technologies, and MediaTek now spans the full stack from compiler-level scheduling to on-chip power state machines. Academic institutions including UESTC are pushing AI-on-PDN convergence at the research frontier.

The five core technical sub-domains are impedance profile shaping, active noise monitoring and filtering, AI-predicted workload-aware power management, dynamic voltage-frequency scaling (DVFS) integrated into accelerator architecture, and package-level and system-level PDN co-optimization. PatSnap Analytics provides patent landscape tooling to map these sub-domains and track assignee activity across jurisdictions.

PatSnap Eureka Dataset spans 2009–2025 across US, CN, IN, and PCT jurisdictions. Represents a snapshot of innovation signals, not a comprehensive industry view. Explore the data ↗
2009
Earliest foundational PDN record in dataset
2025
Latest leading-edge filings (AI-on-PDN)
2019–21
Clear inflection point driven by DNN scaling and sub-10 nm nodes
79
Low-power AI accelerator designs surveyed 2019–2022 (CNNs & DNNs dominant)
Key Technology Approaches

Four Patent Clusters Shaping Low-Noise PDN Innovation

From hierarchical impedance simulation to AI-embedded PDN chips, the patent landscape reveals four distinct engineering strategies for managing supply noise in AI accelerator hardware.

Cluster 1

Hierarchical Impedance Simulation & Pre-Silicon PDN Optimisation

This approach treats the entire hardware stack — chip, package substrate, and motherboard — as a coupled PDN system. Engineers simulate impedance as a function of frequency, identify resonance peaks, and iteratively optimise decoupling capacitor placement, via density, and plane geometry. Haiguang’s three-tier simulation methodology (2025, CN) injects noise at the frequency of maximum impedance across all three stack layers simultaneously, exposing resonance effects that single-tier simulation misses. The Hengxin Dayou DC-DC + LDO combination architecture (2019, CN) establishes a switching regulator stepping voltage to slightly above load requirement, with a low-dropout linear regulator providing a clean final rail that attenuates switching harmonics. Post-silicon correction of PDN noise is extremely difficult; the impedance profile must be shaped correctly before tape-out. Learn more about IP analytics for semiconductor PDN landscapes.

Pre-silicon · Three-tier stack · Resonance detection
Cluster 2

Active & AI-Driven Noise Monitoring with Closed-Loop PDN Control

Rather than relying solely on passive filtering, this cluster deploys on-chip sensors and controllers that monitor supply voltage in real time and dynamically adjust active filtering elements to maintain noise within a target budget. Amazon’s Intelligent Power Noise Reduction patent (2025, US) monitors both voltage noise level and bit error rate simultaneously — the first retrieved example of BER-coupled PDN control specifically for AI workloads. UESTC’s Full-Chip Low-Power Supply Network Chip (2026 priority, CN) embeds a dedicated AI computation module on the PDN chip itself, comprising a time-series feature extractor, a power consumption prediction engine, and a decision logic unit driving a dynamic adjustment unit with a switch matrix and voltage trimmer. Shanghai Hailuchuang’s high-PSRR PLL circuit (2025, CN) uses a dedicated LDO to supply the VCO with a clean, isolated rail, dynamically optimising VCO operating voltage to match instantaneous load.

BER-coupled control · AI prediction · Closed-loop
Cluster 3

DVFS & Workload-Aware Power State Management Co-Designed with Accelerator Architecture

This cluster integrates voltage and frequency management directly into the accelerator control fabric, so that compute intensity, memory traffic, and thermal state are used as inputs to PDN control decisions. NVIDIA’s Optimal Operating Point Estimator (2021, US) determines optimal operating states and voltage state combinations that balance performance against power consumption for chips with multiple hardware components sharing a power/thermal budget. MediaTek’s DNN traffic-aware power management (2024, US) uses compiler-generated information about DNN compute loading and NoC traffic to select between routing schemes, enabling the power manager to anticipate traffic spikes and pre-condition the PDN before the current surge arrives. Graphic Era University’s energy-efficient accelerator device (2025, IN) uses an intelligent workload controller with distributed thermal sensors and cooling pathways.

Compiler-PDN · DVFS · Thermal co-management
Cluster 4

Near-Threshold Voltage (NTV) Supply Design for Ultra-Low-Power Accelerator Operation

NTV operation places the supply voltage at or just above the transistor threshold, maximising energy efficiency but dramatically increasing susceptibility to supply noise. PDN design in this regime requires extremely tight impedance control and noise budgeting. Research across Intel 32, 22, and 14 nm nodes established that reliable NTV operation requires PDN designs validated from nominal voltage down to the threshold voltage. A 2016 study introduced multiple-clock-domain design combined with NTV supply for ANN accelerators, where careful clock domain isolation reduces simultaneous switching noise — a critical PDN noise source. Xi’an University of Posts and Telecommunications’ 2025 coprocessor architecture combines a near-threshold supply system with adaptive body biasing, where the threshold voltage itself is dynamically adjusted to track the supply rail, maintaining noise margin even as the rail varies.

NTV · Adaptive body biasing · Clock domain isolation
PatSnap Eureka Cluster analysis derived from patent and literature records 2009–2025. Assignees include Amazon, NVIDIA, MediaTek, Haiguang, UESTC, and Xi’an University of Posts and Telecommunications. Explore all clusters ↗
Data Visualisation

Geographic Distribution & Innovation Timeline

Patent jurisdiction data and filing timeline reveal where PDN innovation is concentrated and how the field has evolved from foundational converter design to AI-integrated power management.

Patent Assignee Jurisdiction Distribution

US and CN jurisdictions dominate the dataset; US filings concentrate in hyperscaler and GPU/AI chip companies while CN filings are largely from universities and domestic design houses.

Patent Jurisdiction Distribution: US 5 patents, CN 4 patents, IN 1 patent, PCT/WO 1 patent Donut chart showing geographic distribution of patent records in the low-noise PDN for AI accelerator chips dataset. US and CN are the dominant jurisdictions. Source: PatSnap Eureka 2025.

Innovation Timeline: Filing Intensity by Period

A clear inflection point around 2019–2021 reflects the rapid scaling of DNN workloads and the push below 10 nm, with the 2023–2025 cluster dominated by AI-loop integration and system-level simulation.

PDN Innovation Filing Intensity: Foundational pre-2018 3 records, Development cluster 2019-2022 7 records, Leading-edge 2023-2025 8 records Bar chart showing filing intensity across three innovation periods for low-noise PDN AI accelerator chip patents and literature. The 2023–2025 period shows the highest concentration of leading-edge filings. Source: PatSnap Eureka.
PatSnap Eureka Filing intensity counts are approximate and based on the retrieved dataset only; they do not represent the full industry filing volume. Explore the timeline ↗
Application Domains

From Hyperscale Datacentres to Edge and Space

Low-noise PDN techniques identified in this dataset serve four distinct deployment contexts, each with its own noise budget, space constraints, and reliability requirements.

Datacenter AI ASICs
High-Performance Training & Inference
Supply noise manifests as timing violations in high-speed memory interfaces and arithmetic unit metastability. Amazon’s intelligent PDN controller and NVIDIA’s operating point estimator both explicitly target integrated circuit devices executing AI workloads at scale.
Multi-Core AI SoCs with NoC
MediaTek’s DNN traffic-aware power management addresses multi-core AI SoCs where traffic bursts from matrix engines cause simultaneous switching noise events on the Network-on-Chip interconnect.
Edge AI & IoT
Constrained Board Space
Edge deployment constrains both power budget and board space, making PDN design more challenging: less room for bulk decoupling capacitance and less power headroom for linear regulators.
79 Designs Surveyed 2019–2022
A 2022 survey identifies 79 low-power AI accelerator designs between 2019–2022, noting CNNs and DNNs are dominant targets and supply voltage management is a critical design variable. FPGA-based CNN accelerators achieving under 5 W require careful on-board PDN design.
Unlock Space & Scientific PDN Techniques
Discover how radiation-hardened neural inference chips and scientific detector front-end systems inform AI accelerator PDN noise tolerance.
Radiation-hardened PDNsLHC converter methodology+ more
Generate full report →
PatSnap Eureka Application domain analysis drawn from retrieved patent and literature records. Edge AI survey covers 79 designs 2019–2022. Explore edge AI PDN patents ↗
Emerging Directions

Four Frontiers Identified in 2023–2025 Filings

The most recent patent cluster reveals a convergence of AI inference, compiler toolchains, and power delivery hardware into tightly integrated autonomous systems.

AI-on-PDN-Chip Architectures (2025–2026)

UESTC’s patent embeds a dedicated power-prediction neural network directly on the PDN management chip, comprising a time-series feature extractor (sliding window buffer and frequency-domain converter), a power consumption prediction engine, and a decision logic unit. This moves PDN control from reactive (sense-then-adjust) to predictive (forecast-then-pre-condition). PatSnap Analytics can map the white space around this emerging architecture.

BER-Coupled PDN Control (2025)

Amazon’s active noise reduction patent links PDN tuning directly to bit error rate measurements, creating a functional quality metric that feeds back into the supply network. This is a significant departure from voltage-amplitude-only noise budgeting and represents the first retrieved example of BER-coupled PDN control specifically for AI workloads. IP strategists should evaluate freedom-to-operate around this active, sensor-driven PDN control architecture.

Unlock Compiler-PDN Co-Design & Adaptive NTV
Access the full analysis of compiler-PDN interface white space and adaptive body-biased near-threshold supply systems from 2024–2025 filings.
Compiler-PDN interfaceAdaptive body biasingIn-memory compute+ more
Unlock full report →
PatSnap Eureka Emerging direction analysis based on 2023–2025 filings from Amazon, UESTC, MediaTek, and Xi’an University of Posts and Telecommunications. Explore emerging filings ↗
Strategic Implications

What the PDN Patent Landscape Means for R&D and IP Teams

Five actionable signals for engineering teams, IP strategists, and product developers working on AI accelerator power delivery.

Pre-silicon PDN simulation must include all three stack layers. Haiguang’s three-tier (chip/substrate/motherboard) simulation methodology demonstrates that single-layer or two-layer simulation leaves cross-resonance effects undetected, resulting in post-silicon surprises. R&D teams should instrument PDN simulation chains to inject noise at maximum-impedance frequencies across all stack levels simultaneously. Tools from PatSnap Analytics can help identify which assignees hold the strongest positions in this space.

Active PDN controllers are becoming an IP battleground. Amazon’s 2025 patent combining voltage noise sensing and BER monitoring in a unified active filter control loop represents a defensible IP position at the system-system interface. IP strategists entering the AI accelerator space should evaluate freedom-to-operate around active, sensor-driven PDN control architectures. The PatSnap customer success programme includes IP landscape services for exactly this use case.

Compiler-PDN interface is an underexplored white space. MediaTek’s 2024 filing demonstrates that the compiler already has information — in the form of workload scheduling and NoC routing — that can be used to pre-condition the PDN before current surges arrive. Relatively few patents address this interface, representing an opportunity for EDA tool vendors and chip designers to establish early IP positions. Researchers can access raw patent data programmatically via PatSnap Open API.

AI-on-PDN convergence creates new product categories. UESTC’s dedicated AI computation module embedded on the PDN chip itself suggests a path toward autonomous power delivery subsystems that self-optimise without host CPU involvement. Product developers should evaluate whether PDN intelligence should reside on the main die, a companion die, or a dedicated power management IC in the package.

Near-threshold supply design requires co-optimisation with noise immunity at the transistor level. NTV operation radically reduces noise margin; the adaptive body-biasing approach (Xi’an, 2025) and in-memory compute (eliminating large data-movement current transients) are complementary techniques that must be evaluated together with PDN impedance budgets rather than in isolation. The IEEE and JEDEC standards bodies publish relevant supply noise specifications for memory interfaces.

PatSnap Eureka Strategic implications derived from patent assignee analysis and filing cluster data 2021–2025. Explore IP landscape ↗
Key Assignees in Dataset
  • Amazon Technologies, Inc. — BER-coupled active PDN control (US, 2025)
  • NVIDIA Corporation — Optimal operating point estimator (US, 2021)
  • MediaTek Inc. — DNN traffic-aware power management (US, 2024)
  • Haiguang (Chengdu Hygon) — Three-tier PDN simulation (CN, 2025)
  • UESTC — AI-on-PDN chip architecture (CN, 2026 priority)
  • Xi’an Posts & Telecom. — Adaptive body-bias NTV supply (CN, 2025)
  • Shanghai Hailuchuang — High-PSRR PLL with isolated VCO rail (CN, 2025)
  • Graphic Era University — Energy-efficient DVFS accelerator (IN, 2025)
White Space Opportunities
  • Compiler-to-PDN controller interface patents
  • Multi-die package PDN co-optimisation (chiplet architectures)
  • BER-feedback PDN control outside Amazon’s claim scope
  • In-memory compute + NTV supply co-design for edge AI
Frequently asked questions

Low-Noise PDN Design for AI Accelerators — Key Questions Answered

Still have questions? PatSnap Eureka can answer them instantly from patent and research data. Ask Eureka ↗
PatSnap Eureka

Generate Your Own PDN Patent Landscape Report

Join 18,000+ innovators using PatSnap Eureka to generate reports like this one for any technology area — from impedance simulation to AI-driven power management.

Ask anything about low-noise PDN design for AI accelerator chips.
PatSnap Eureka searches patents and research literature to answer instantly.
Powered by PatSnap Eureka
Link copied to clipboard