ML ASIC Instruction-Rate Throttling for Overcurrent Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The rapid evolution of machine learning applications necessitates short design and manufacturing cycles for ASICs, leading to potential overheating and damage due to power demands exceeding system capacity, resulting in shutdowns and loss of computation.

Innovation Solution

A system and method involving a voltage regulator that monitors current load, detects overcurrent conditions, and sends signals to an integrated circuit (IC) die to throttle processing rates using a throttling mask, synchronized with the IC's internal clock, to manage power usage and prevent overheating.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If the ASIC operates at high power to meet evolving ML application demands, then computational productivity is improved, but the risk of overheating and damage increases

Engineering Contradiction:
Improvecomputational productivityVSAvoidoverheating risk
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The system continuously monitors the actual power consumption of the ASIC and uses this feedback to dynamically adjust the instruction rate. The power management circuit receives power consumption information and modifies the throttling mask accordingly, creating a closed-loop control system that balances computational productivity with thermal safety.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The invention implements dynamic power management by allowing the ASIC's instruction rate to vary in real-time based on actual power consumption conditions. The throttling mask is continuously updated to reflect current operational state, enabling the system to adapt between high-performance mode and power-saving mode as needed.

Inventive Principle:
Principle #15Dynamics

2Reliability

If the power supply ceases power delivery when overcurrent threshold is exceeded, then system safety is improved, but computation continuity is lost

Engineering Contradiction:
Improvesystem safetyVSAvoidcomputation time loss
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The invention introduces an intermediary throttling mechanism between the power supply and the ASIC. Instead of the power supply directly shutting down the ASIC when overcurrent occurs, the power management circuit acts as a mediator that throttles the instruction rate to prevent overcurrent conditions, thereby maintaining computation continuity while ensuring safety.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system takes preliminary action by proactively throttling the instruction rate before overcurrent conditions can cause damage. The power management circuit monitors power consumption and adjusts the throttling mask in advance to prevent reaching dangerous current levels, avoiding the need for shutdowns.

Inventive Principle:
Principle #10Preliminary action

3Strength

If the ASIC is shut down to prevent overheating, then hardware protection is improved, but computational availability decreases

Engineering Contradiction:
Improvehardware protectionVSAvoidcomputational availability
Core Design Contradiction:
StrengthVSProductivity

Solution Approach 1:

The system dynamically adjusts the ASIC's operational state by modifying the instruction rate through throttling masks rather than complete shutdowns. This dynamic control allows the ASIC to operate at reduced capacity when approaching thermal limits, maintaining computational availability while providing hardware protection.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The invention changes the operational parameters of the ASIC by adjusting the instruction rate based on power consumption conditions. The power management circuit modifies the throttling mask to change the ASIC's processing speed, allowing continuous operation at varying performance levels rather than binary on/off states.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3879383B1Power throttling mechanism using instruction rate limiting in high power machine-learning asics
Publication Date: 2026.03.11 GOOGLE LLC
  • EP3879383B1 patent drawingFigure 1A
  • EP3879383B1 patent drawingFigure 1B
  • EP3879383B1 patent drawingFigure 2

AI summary

A system contains a machine learning application specific integrated circuit (ASIC) and a power supply unit. The power supply unit and the ASIC are configured to be in data communication through dedicated pins on the ASIC and the power supply unit. The power supply unit detects a present power consumption of the ASIC. Upon determining that a threshold condition has been met, the power supply unit, responsive to the condition sends a digital signal to the ASIC. The ASIC contains a synchronizer which synchronizes the digital signal to be consistent with the ASICs internal clock frequency. A chip manager the synchronized signal and other signals to generate a throttling mask. The throttling mask is sent to a sequencer of the ASIC, which then limits the instruction flow into the processing units of the ASIC based on the mask. This in turn limits the power being consumed by the ASIC.