0% found this document useful (0 votes)
8 views14 pages

Low-Power Clock Design for NPUs

The document presents a project on low-power clock architecture for Neural Processing Units (NPUs), detailing the importance of clock signals in synchronizing operations and enhancing performance. It discusses various components of the clock module, including crystal oscillators, PLLs, and clock gating, along with proposed designs for optimizing power consumption and performance. The presentation also compares different PLL architectures and outlines a proposed design combining ADPLL and ILO for improved efficiency.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
8 views14 pages

Low-Power Clock Design for NPUs

The document presents a project on low-power clock architecture for Neural Processing Units (NPUs), detailing the importance of clock signals in synchronizing operations and enhancing performance. It discusses various components of the clock module, including crystal oscillators, PLLs, and clock gating, along with proposed designs for optimizing power consumption and performance. The presentation also compares different PLL architectures and outlines a proposed design combining ADPLL and ILO for improved efficiency.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

Low-Power Clock Architecture for NPU:

Design & Optimization


Presented by
Rakesh Arumalla (221206), (3rd Year)
Vikas Naik (221233), (3rd Year)
Tharun Nayak (221254), (3rd Year)

Project Supervisor: Prof. G. Seetharaman, C2S Project Investigator


Department of Electronics and Communication Engineering
Indian Institute of Information Technology Tiruchirappalli
Date: 24/02/2025
Overview

 Introduction

 Literature Review

 Proposed Research

 References

03/03/2025 C2S Project Presentation 2


Why is Clock Needed for NPUs?

Clock Signal Importance:


• Synchronizes all operations within the NPU

• Ensures data processing occurs in a controlled manner

• Facilitates parallel execution of AI computations

Impact of Clock on Performance:


• Higher clock speeds enable faster neural network processing

• Stability in clock signal reduces computational errors

• Power-efficient clocking extends battery life in embedded AI application

03/03/2025 C2S Project Presentation 3


Parts of the Clock Module
Crystal Oscillator:

• Provides a stable frequency reference

• Determines the base clock signal for the entire system

Phase-Locked Loop (PLL):

• Multiplies and stabilizes the clock frequency

• Reduces phase noise and jitter

Clock Dividers:

• Generate different clock signals from a single source

• Allow multiple clock domains in NPUs


03/03/2025 C2S Project Presentation
Parts of the Clock Module

Clock Gating:
• Reduces power consumption by disabling unused clock signals

Dynamic Frequency Scaling (DFS):


• Adjusts clock speed based on processing demand to optimize
power

Clock Monitors:
• Ensure stability, detect anomalies, and maintain timing accuracy

Clock Distribution

03/03/2025 C2S Project Presentation 5


Proposed Crystal Frequency

Crystal Frequency Use Case Example Chips


Qualcomm Snapdragon NPU,
19.2 MHz Mobile SoCs, Edge AI
Google Edge TPU
NVIDIA Jetson AI Chips, ARM AI
24 MHz Consumer AI Chips, GPUs
Cores
25 MHz Datacenter NPUs, Networking Broadcom AI Switches, Google TPU
Huawei Kirin NPU, MediaTek AI
26 MHz 5G AI, Embedded AI
Chips
40 MHz High-Speed AI Tesla Dojo, IBM AI Chips
Custom Edge AI NPUs, High-
50 MHz Custom AI ASICs
Performance AI
NVIDIA Tensor Cores, AMD Instinct
100 MHz Server AI, GPUs
AI
125 MHz High-Speed Compute Intel Habana Gaudi AI Accelerator

03/03/2025 C2S Project Presentation 6


PLL-Architecture

03/03/2025 C2S Project Presentation 7


Types of Architectures
PLL Type Power Consumption Jitter Performance Complexity Best Use Case
General-purpose,
Charge-Pump PLL (CP-PLL) Medium Low jitter High microcontrollers, mixed-
signal SoCs

Ultra-low-power IoT, Edge


All-Digital PLL (ADPLL) Very Low Medium High
AI, processors
Power-sensitive
Digital PLL (DPLL) Low Medium Medium
applications, IoT
RF circuits, energy-
Injection-Locked PLL (ILPLL) Extremely Low High jitter Low
harvesting systems

Ultra-low power, clock


Sub-Sampling PLL (SSPLL) Low Very Low (Best) High
generators for ADCs

Fractional-N PLL Medium Low jitter High RF transceivers, telecom

Hybrid PLL (Analog + High-speed, low-power


Low Very Low Medium
Digital) mixed-signal designs

03/03/2025
C2S Project Presentation 8
Differences in Traditional PLL and ILO-PLL

03/03/2025 C2S Project Presentation 9


Proposed Design for PLL
ADPLL + ILO (All-Digital PLL + Injection-Locked Oscillator)

03/03/2025 C2S Project Presentation 10


Differences between Two stage Design and Single Stage-
(MOD-30)

Factor ILO + ADPLL (5×6) ✅ ILO + Mod-30 Counter ❌

Switching Activity Lower ✅ (Optimized loop operation) Higher ❌ (Frequent toggling)

Leakage Power Lower ✅ (Optimized transistors) Higher ❌ (Large counter logic)


Jitter Performance Better ✅ Higher Jitter ❌
Clock Stability Higher Stability ✅ More Noise ❌

03/03/2025 C2S Project Presentation 11


Components in ADPLL(All-digital phase-locked loop)

03/03/2025 C2S Project Presentation 12


Proposed Design Outline

03/03/2025 C2S Project Presentation 13


References
 List 5 to 10 References
 In IEEE Reference format
 List in Alphabetical order

Example:
1. G. Eason, B. Noble, and I. N. Sneddon, “On certain integrals of Lipschitz-Hankel type involving products of Bessel
functions,” Phil. Trans. Roy. Soc. London, vol. A247, pp. 529–551, April 1955.

Author Name, Author Name, “Title of the Paper”, Journal Name, Vol. No: xy, Issue No: xy, pp. xyz-xyz , Month
Year.

Thank You

03/03/2025 C2S Project Presentation 14

You might also like